VampirTrace 5.14.4 User Manual


TU Dresden
Center for Information Services and
High Performance Computing (ZIH)
01062 Dresden
Germany

http://www.tu-dresden.de/zih
http://www.tu-dresden.de/zih/vampirtrace

Contact: vampirsupport@zih.tu-dresden.de


Contents

This documentation describes how to apply VampirTrace to an application in order to generate trace files at execution time. This step is called instrumentation. It furthermore explains how to control the runtime measurement system during execution (tracing). This also includes performance counter sampling as well as selective filtering and grouping of functions.

Introduction

VampirTrace consists of a tool set and a runtime library for instrumentation and tracing of software applications. It is particularly tailored to parallel and distributed High Performance Computing (HPC) applications.

The instrumentation part modifies a given application in order to inject additional measurement calls during runtime. The tracing part provides the actual measurement functionality used by the instrumentation calls. By this means, a variety of detailed performance properties can be collected and recorded during runtime. This includes function enter and leave events, MPI communication, OpenMP events, and performance counters.

After a successful tracing run, VampirTrace writes all collected data to a trace file in the Open Trace Format (OTF)[*]. As a result, the information is available for post-mortem analysis and visualization by various tools. Most notably, VampirTrace provides the input data for the Vampir analysis and visualization tool[*].

VampirTrace is included in OpenMPI 1.3 and later versions. If not disabled explicitly, VampirTrace is built automatically when installing OpenMPI[*].

Trace files can quickly become very large, especially with automatic instrumentation. Tracing applications for only a few seconds can result in trace files of several hundred megabytes. To protect users from creating trace files of several gigabytes, the default behavior of VampirTrace limits the internal buffer to 32 MB per process. Thus, even for larger scale runs the total trace file size will be moderate. Please read Section 3.3 on how to remove or change this limit.

VampirTrace supports various Unix and Linux platforms that are common in HPC nowadays. It is available as open source software under a BSD License.

The following list shows a summary of all instrumentation and tracing features that VampirTrace offers. Note that not all features are supported on all platforms.



Tracing of user functions ⇒ Chapter 2

MPI Tracing ⇒ Chapter 2

OpenMP Tracing ⇒ Chapter 2

Pthread Tracing

Java Tracing ⇒ Section 2.8

3rd-Party Library tracing ⇒ Section 2.9

MPI Correctness Checking ⇒ Section 4.10

User API

Performance Counters ⇒ Sections 4.1 and 4.2

CPU ID Tracing ⇒ Section 4.4

Child Process Execution Tracing ⇒ Section 4.9

I/O Tracing ⇒ Section 4.8

Memory Allocation Tracing ⇒ Section 4.3

Filtering & Grouping ⇒ Chapter 5

OTF Output ⇒ Chapter 3


Instrumentation

To perform measurements with VampirTrace, the user's application program needs to be instrumented, i.e., at specific points of interest (called ``events'') VampirTrace measurement calls have to be activated. As an example, common events are, amongst others, entering and leaving of functions as well as sending and receiving of MPI messages.

VampirTrace handles this automatically by default. In order to enable the instrumentation of function calls, the user only needs to replace the compiler and linker commands with VampirTrace's wrappers, see Section 2.1 below. VampirTrace supports different ways of instrumentation as described in Section 2.2.


Compiler Wrappers

All the necessary instrumentation of user functions, MPI, and OpenMP events is handled by VampirTrace's compiler wrappers (vtcc, vtcxx, and vtfort). In the script used to build the application (e.g. a makefile), all compile and link commands should be replaced by the VampirTrace compiler wrapper. The wrappers perform the necessary instrumentation of the program and link the suitable VampirTrace library. Note that the VampirTrace version included in OpenMPI 1.3 has additional wrappers (mpicc-vt, mpicxx-vt, mpif77-vt, and mpif90-vt) which are like the ordinary MPI compiler wrappers (mpicc, mpicxx, mpif77, and mpif90) with the extension of automatic instrumentation.

The following list shows some examples specific to the parallelization type of the program:

The VampirTrace compiler wrappers automatically try to detect which parallelization method is used by means of the compiler flags (e.g. -lmpi, -openmp or -pthread) and the compiler command (e.g. mpif90). If the compiler wrapper failed to detect this correctly, the instrumentation could be incomplete and an unsuitable VampirTrace library would be linked to the binary. In this case, you should tell the compiler wrapper which parallelization method your program uses by using the switches -vt:mpi, -vt:mt, and -vt:hyb for MPI, multithreaded, and hybrid programs, respectively. Note that these switches do not change the underlying compiler or compiler flags. Use the option -vt:verbose to see the command line that the compiler wrapper executes. The default settings of the compiler wrappers can be modified in the files share/vampirtrace/vtcc-wrapper-data.txt (and similar for the other languages) in the installation directory of VampirTrace. The settings include compilers, compiler flags, libraries, and instrumentation types. You could for instance modify the default C compiler from gcc to mpicc by changing the line compiler=gcc to compiler=mpicc. This may be convenient if you instrument MPI parallel programs only.


Instrumentation Types

The wrapper option -vt:inst <insttype> specifies the instrumentation type to be used. The following values for <insttype> are possible:

To determine which instrumentation type will be used by default and which instrumentation types are available on your system have a look at the entry inst_avail in the wrapper's configuration file (e.g. share/vampirtrace/vtcc-wrapper-data.txt in the installation directory of VampirTrace for the C compiler wrapper).

Type vtcc -vt:help for otheroptions that can be passed to VampirTrace's compiler wrapper.


Automatic Instrumentation

Automatic instrumentation is the most convenient method to instrument your program. If available, simply use the compiler wrappers without any parameters, e.g.:



% vtfort hello.f90 -o hello

Supported Compilers

VampirTrace supports following compilers for automatic instrumentation:


Notes for Using the GNU, Intel, PathScale, or Open64 Compiler

For these compilers the command nm is required to get symbol information of the running application executable. For example on Linux systems, this program is a part of the GNU Binutils, which is downloadable from http://www.gnu.org/software/binutils. To get the application executable for nm during runtime, VampirTrace uses the /proc file system. As /proc is not present on all operating systems, automatic symbol information might not be available. In this case, it is necessary to set the environment variable VT_APPPATH to the pathname of the application executable to get symbols resolved via nm.

Should any problems emerge to get symbol information automatically, then the environment variable VT_GNU_NMFILE can be set to a symbol list file, which is created with the command nm, like:



% nm hello > hello.nm


To get the source code line for the application functions use nm -l on Linux systems. VampirTrace will include this information into the trace. Note that the output format of nm must be written in BSD-style. See the manual page of nm to obtain help for dealing with the output format setting.


Notes on Instrumentation of Inline Functions

Compilers behave differently when they automatically instrument inlined functions. The GNU and Intel ≥10.0 compilers instrument all functions by default when they are used with VampirTrace. They therefore switch off inlining completely, disregarding the optimization level chosen. One can prevent these particular functions from being instrumented by appending the following attribute to function declarations, hence making them able to be inlined (this works only for C/C++):



__attribute__ ((__no_instrument_function__))

The PGI and IBM compilers prefer inlining over instrumentation when compiling with enabled inlining. Thus, one needs to disable inlining to enable the instrumentation of inline functions and vice versa.

The bottom line is that a function cannot be inlined and instrumented at the same time. For more information on how to inline functions read your compiler's manual.


Instrumentation of Loops with OpenUH Compiler

The OpenUH compiler provides the possibility of instrumenting loops in addition to functions. To use this functionality add the compiler flag -OPT:instr_loop. In this case loops induce additional events including the type of loop (e.g. for, while, or do) and the source code location.


Manual Instrumentation

Using the VampirTrace API

The VT_USER_START, VT_USER_END calls can be used to instrument any user-defined sequence of statements. 

Fortran: 
           #include "vt_user.inc"
           VT_USER_START('name')
           ...
           VT_USER_END('name')


C:
           #include "vt_user.h"
           VT_USER_START("name");
           ...
           VT_USER_END("name");
If a block has several exit points (as it is often the case for functions), all exit points have to be instrumented with VT_USER_END, too.

For C++ it is simpler as is demonstrated in the following example. Only entry points into a scope need to be marked. The exit points are detected automatically when C++ deletes scope-local variables. 

C++:
           #include "vt_user.h"
           {
             VT_TRACER("name");
             ...
           }

The instrumented sources have to be compiled with -DVTRACE for all three languages, otherwise the VT_* calls are ignored. Note that Fortran source files instrumented this way have to be preprocessed, too.

In addition, you can combine this particular instrumentation type with all other types. In such a way, all user functions can be instrumented by a compiler while special source code regions (e.g. loops) can be instrumented by VT's API.

Use VT's compiler wrapper (described above) for compiling and linking the instrumented source code, such as:

Note that you can also use the option -vt:inst manual with non-instrumented sources. Binaries created in this manner only contain MPI and OpenMP instrumentation, which might be desirable in some cases.


Measurement Controls


Switching tracing on/off:

In addition to instrumenting arbitrary blocks of code, one can use the VT_ON/ VT_OFF instrumentation calls to start and stop the recording of events. These constructs can be used to stop recording of events for a part of the application and later resume recording. For example, as is demonstrated in the following C/C++ code snippet, one could not collect trace events during the initialization phase of an application and turn on tracing for the computation part. 
           int main() {
             ...
             VT_OFF();
             initialize();
             VT_ON();
             compute();
             ...
           }
Furthermore the "on/off" functionality can be used to control the tracing behavior of VampirTrace and allows to trace only parts of interests. Therefore the amount of trace data can be reduced essentially. To check whether if tracing is enabled or not use the call VT_IS_ON.

Trace buffer rewind:

An alternative to the "on/off" functionality is the buffer rewind approach. It is useful when the program should decide dynamically after a specific code section (i.e. a time step or iteration) if this section has been interesting (i.e. anomalous/slow behavior) and should be recorded to the trace file. The key difference to "on/off" is that you do not need to know a priori if a section should be recorded.

Use the instrumentation call VT_SET_REWIND_MARK at the beginning of a (possibly not interesting) code section. Later, you can decide to rewind the trace buffer to the mark with the call VT_REWIND. All recorded trace data between the mark and the rewind call will be dropped. Note, that only one mark can be set at a time. The last call to VT_SET_REWIND_MARK will be considered when rewinding the trace buffer. This simplified Fortran code example sketches how the rewind approach can be used: 

           do step=1,number_of_time_steps
             VT_SET_REWIND_MARK()
             call compute_time_step(step)
             if(finished_as_expected) VT_REWIND()
           end do

Intermediate buffer flush:

In addition to an automated buffer flush when the buffer is filled, it is possible to flush the buffer at any point of the application. This way you can guarantee that after a manual buffer flush there will be a sequence of the program with no automatic buffer flush interrupting. To flush the buffer you can use the call VT_BUFFER_FLUSH.

Intermediate time synchronisation:

VampirTrace provides several mechanisms for timer synchronization (⇒ Section 3.7). In addition it is also possible to initiate a timer synchronization at any point of the application by calling VT_TIMESYNC. Please note that the user has to ensure that all processes are actual at a synchronized point in the program (e.g. at a barrier). To use this call make sure that the enhanced timer synchronization is activated (set the environment variable VT_ETIMESYNC ⇒ Section 3.2).

Intermediate counter update:

VampirTrace provides the functionality to collect the values of arbitrary hardware counters. Chosen counter values are automatically recorded whenever an event occurs. Sometimes (e.g. within a long-lasting function) it is desirable to get the counter values at an arbitrary point within the program. To record the counter values at any given point you can call VT_UPDATE_COUNTER.

Note:

For all three languages the instrumented sources have to be compiled with -DVTRACE. Otherwise the VT_* calls are ignored.
In addition, if the sources contains further VampirTrace API calls and only the calls for measurement controls shall be disabled, then the sources have to be compiled with -DVTRACE_NO_CONTROL, too.


Source Instrumentation Using PDT/TAU

TAU instrumentation combines the advantages of compiler and manual instrumentation and has further advantages. Like compiler instrumentation it works automatically, like on manual instrumentation you have a filtered set of events, this is especially recommended for C++, because STL-constructor calls are suppressed. Unlike with compiler instrumentation you get an optimized binary - this solves the issue described in Section 2.3.3. In the simpliest case you just run the compiler wrappers with -vt:inst tauinst option:



% vtcc -vt:inst tauinst hello.c -o hello

Requirements for TAU instrumentation:

To work with TAU instrumenation you need the Program Database Toolkit. You have to make sure, to have cparse and tau_instrumentor in your $PATH. The PDToolkit can be downloaded from http://www.cs.uoregon.edu/research/pdt/home.php.

Include/Exclude Lists:

tau_instrumentor provides a mechanism to include and exclude files or functions from instrumenation. The lists are deposed
in a single file, that is announced to tau_instrumentor via the option
-f <filename>. This file contains up to four lists which begin with
BEGIN[_FILE]_<INCLUDE|EXCLUDE>_LIST. The names in between may contain wildcards as ``?'', ``*', and ``#'', each entry gets a new line. The lists end with END[_FILE]_<INCLUDE|EXCLUDE>_LIST. For further information on selective profiling have a look at the TAU documentation[*]. To announce the file through the compiler wrapper use the option -vt:tau: 
   % vtcc -vt:inst tauinst hello.c -o hello \
     -vt:tau '-f <filename>'


Binary Instrumentation Using Dyninst

The option -vt:inst dyninst is used with the compiler wrapper to instrument the application during runtime (binary instrumentation), by using Dyninst[*]. Recompiling is not necessary for this kind of instrumentation, but relinking:



% vtfort -vt:inst dyninst hello.o -o hello


The compiler wrapper dynamically links the library libvt-dynatt.so to the application. This library attaches the mutator-program vtdyn during runtime which invokes the instrumentation by using Dyninst.

To prevent certain functions from being instrumented you can use the runtime function filtering as explained in Section 5.1. All additional overhead, due to instrumentation of these functions, will be removed.

VampirTrace also allows binary instrumentation of functions located in shared libraries. For this to work a colon-separated list of shared library names has to be given in the environment variable VT_DYN_SHLIBS:



VT_DYN_SHLIBS=libsupport.so:libmath.so


Static Binary Instrumentation

In order to avoid the overhead introduced by Dyninst during runtime, the tool vtdyn can be used for binary instrumentation before application launch. To accomplish this, the -o or -output switch can be used to specify the output binary. Note that the application must be linked to the corresponding VampirTrace library.

Example

To apply binary instrumentation to the executable a.out the following command is nescessary:



% vtdyn -o dyninst_a.out ./a.out


Runtime Instrumentation Using VTRun

Besides the already described instrumentation at compile-time, VampirTrace also supports runtime instrumention using the vtrun command. Prepending the actual call to the application will transparently add instrumentation support and launch the application. This includes support function instrumentation by Dyninst (Section 2.6) as well as MPI communication tracing. In order to enable instrumentation for user functions the user has to specify the -dyninst command line switch.

Example

In order to add tracing support to an already existing executable, only a small change to the startup command has to be made. Assuming the usual way of calling the application looks like:



% mpirun -np 4 ./a.out

By putting the call to vtrun directly before the actual application call, instrumention support will be enabled at runtime:



% mpirun -np 4 vtrun ./a.out

For more information about the tool vtrun see Section B.6.


Tracing Java Applications Using JVMTI

In addition to C, C++, and Fortran, VampirTrace is capable of tracing Java applications. This is accomplished by means of the Java Virtual Machine Tool Interface (JVMTI) which is part of JDK versions 5 and later. If VampirTrace was built with Java tracing support, the library libvt-java.so can be used as follows to trace any Java program:



% java -agentlib:vt-java ...

Or more easier, by replacing the usal Java application launcher java by the command vtjava:



% vtjava ...

When tracing Java applications, you probably want to filter out dispensable function calls. Please have a look at Sections 5.1 and 5.2 to learn about different ways for excluding parts of the application from tracing.


Tracing Calls to 3rd-Party Libraries

VampirTrace is also capable to trace calls to third party libraries, which come with at least one C header file even without the library's source code. If VampirTrace was built with support for library tracing (the CTool library[*] is required), the tool vtlibwrapgen can be used to generate a wrapper library to intercept each call to the actual library functions. This wrapper library can be linked to the application or used in combination with the LD_PRELOAD mechanism provided by Linux. The generation of a wrapper library is done using the vtlibwrapgen command and consists of two steps. The first step generates a C source file, providing the wrapped functions of the library header file:



% vtlibwrapgen -g SDL -o SDLwrap.c /usr/include/SDL/*.h

This generates the source file SDLwrap.c that contains wrapper-functions for all library functions found in the header-files located in /usr/include/SDL/ and instructs VampirTrace to assign these functions to the new group SDL.

The generated wrapper source file can be edited in order to add manual instrumentation or alter attributes of the library wrapper. A detailed description can be found in the generated source file or in the header file vt_libwrap.h which can be found in the include directory of VampirTrace.

To adapt the library instrumentation it is possible to pass a filter file to the generation process. The rules are like these for normal VampirTrace instrumentation (see Section 5.1), where only 0 (exclude functions) and -1 (generally include functions) are allowed.

The second step is to compile the generated source file:



% vtlibwrapgen --build --shared -o libSDLwrap SDLwrap.c

This builds the shared library libSDLwrap.so which can be linked to the application or preloaded by using the environment variable LD_PRELOAD:



% LD_PRELOAD=$PWD/libSDLwrap.so <executable>


Runtime Measurement

Running a VampirTrace instrumented application should normally result in an OTF trace file in the current working directory where the application was executed. If a problem occurs, set the environment variable VT_VERBOSE to 2 before executing the instrumented application in order to see control messages of the VampirTrace runtime system which might help tracking down the problem.

The internal buffer of VampirTrace is limited to 32 MB per process. Use the environment variables VT_BUFFER_SIZE and VT_MAX_FLUSHES to increase this limit. Section 3.3 contains further information on how to influence trace file size.


Trace File Name and Location

The default name of the trace file depends on the operating system where the application is run. On Linux, MacOS and Sun Solaris the trace file will be named like the application, e.g. hello.otf for the executable hello. For other systems, the default name is a.otf. Optionally, the trace file name can be defined manually by setting the environment variable VT_FILE_PREFIX to the desired name. The suffix .otf will be added automatically.

To prevent overwriting of trace files by repetitive program runs, one can enable unique trace file naming by setting VT_FILE_UNIQUE to yes. In this case, VampirTrace adds a unique number to the file names as soon as a second trace file with the same name is created. A *.lock file is used to count up the number of trace files in a directory. Be aware that VampirTrace potentially overwrites an existing trace file if you delete this lock file. The default value of VT_FILE_UNIQUE is no. You can also set this variable to a number greater than zero, which will be added to the trace file name. This way you can manually control the unique file naming.

The default location of the final trace file is the working directory at application start time. If the trace file shall be stored in another place, use VT_PFORM_GDIR as described in Section 3.2 to change the location of the trace file.


Environment Variables

The following environment variables can be used to control the measurement of a VampirTrace instrumented executable:

Variable Purpose Default

Global Settings
VT_APPPATH Path to the application executable.
⇒ Section 2.3.2		 -
VT_BUFFER_SIZE Size of internal event trace buffer per process. This is the place where event records are stored, before being written to OTF.
⇒ Section 3.3		 32M
VT_CLEAN Remove temporary trace files? yes
VT_COMPRESSION Write compressed trace files? yes
VT_COMPRESSION_BSIZE Size of the compression buffer in OTF. OTF default
VT_FILE_PREFIX Prefix used for trace filenames. ⇒Sect.3.1
VT_FILE_UNIQUE Enable unique trace file naming? Set to yes, no, or a numerical ID.
⇒ Section 3.1		 no
VT_MAX_FLUSHES Maximum number of buffer flushes.
⇒ Section 3.3		 1
VT_MAX_SNAPSHOTS Maximum number of snapshots to generate. 1024
VT_MAX_THREADS Maximum number of threads per process that VampirTrace reserves resources for. 65536
VT_OTF_BUFFER_SIZE Size of internal OTF buffer. This buffer contains OTF-encoded trace data that is written to file at once. OTF default
VT_PFORM_GDIR Name of global directory to store final trace file in. ./
VT_PFORM_LDIR Name of node-local directory which can be used to store temporary trace files. /tmp/
VT_SNAPSHOTS Enable snapshot generation? Allows Vampir to load subsets of the resulting trace. yes
VT_THREAD_BUFFER_SIZE Size of internal event trace buffer per thread. If not defined, the size is set to 10% of VT_BUFFER_SIZE.
⇒ Section 3.3		 0
VT_UNIFY Unify local trace files afterwards? yes
VT_VERBOSE Level of VampirTrace related information messages: Quiet (0), Critical (1), Information (2) 1

I/O Forwarding (IOFSL)
VT_IOFSL_ASYNC_IO Enable buffered IOFSL writes?
⇒ Section D.4.2		 no
VT_IOFSL_SERVERS Comma-separated list of IOFSL server addresses.
⇒ Section D.4.2		 -
VT_IOFSL_MODE Mode of the IOFSL communication:
(MULTIFILE_SPLIT or MULTIFILE)
⇒ Section D.4.2		 MULTIFILE_SPLIT

Optional Features
VT_CPUIDTRACE Enable tracing of core ID of a CPU?
⇒ Section 4.4		 no
VT_ETIMESYNC Enable enhanced timer synchronization?
⇒ Section 3.7		 no
VT_ETIMESYNC_INTV Interval between two successive synchronization phases in s. 120
VT_GPUTRACE Comma-separated list of GPU tracing options.
⇒ Section 4.5		 no
VT_IOLIB_PATHNAME Provides an alternative library to use for LIBC I/O calls. ⇒ Section 4.8 -
VT_IOTRACE Enable tracing of application I/O calls?
⇒ Section 4.8		 no
VT_IOTRACE_EXTENDED Enable tracing of additional function argument for application I/O calls?
⇒ Section 4.8		 no
VT_EXECTRACE Enable tracing of function calls for creating and controling child processes?
⇒ Section 4.9		 yes
VT_MEMTRACE Enable memory allocation counter?
⇒ Section 4.3		 no
VT_MODE Colon-separated list of VampirTrace modes: Tracing (TRACE), Profiling (STAT).
⇒ Section 3.4		 TRACE
VT_MPICHECK Enable MPI correctness checking via UniMCI? no
VT_MPICHECK_ERREXIT Force trace write and application exit if an MPI usage error is detected? no
VT_MPITRACE Enable tracing of MPI events? yes
VT_MPI_IGNORE_FILTER Enable tracing of MPI communication events although its corresponding functions are filtered? no
VT_OMPTRACE Enable tracing of OpenMP events instrumented by OPARI? yes
VT_PTHREAD_REUSE Reuse IDs of terminated Pthreads? yes
VT_STAT_INTV Length of interval in ms for writing the next profiling record 0
VT_STAT_PROPS Colon-separated list of event types that shall be recorded in profiling mode: Functions (FUNC), Messages (MSG), Collective Ops. (COLLOP) or all of them (ALL)
⇒ Section 3.4		 ALL
VT_SYNC_FLUSH Enable synchronized buffer flush?
⇒ Section 3.6		 no
VT_SYNC_FLUSH_LEVEL Minimum buffer fill level for synchronized buffer flush in percent. 80

Counters
VT_CUPTI_METRICS Specify CUDA hardware counter metrics (CUPTI events) to be recorded with trace events as a colon/VT_METRICS_SEP-separated list of names.
⇒ Section 4.5		 -
VT_CUPTI_EVENTS_SAMPLING Sample CUDA hardware counters during the execution of a kernel.
⇒ Section 4.5		 no
VT_METRICS Specify counter metrics to be recorded with trace events as a colon/VT_METRICS_SEP-separated list of names.
⇒ Section 4.1		 -
VT_METRICS_SEP Separator string between counter specifications in VT_METRICS. :
VT_PLUGIN_CNTR_METRICS Colon-separated list of plugin counter metrics which shall be recorded.
⇒ Section 4.7		 -
VT_RUSAGE Colon-separated list of resource usage counters which shall be recorded.
⇒ Section 4.2		 -
VT_RUSAGE_INTV Sample interval for recording resource usage counters in ms. 100

Binary Instrumentation (Dyninst)
VT_DYN_DETACH Detach Dyninst mutator-program vtdyn from application process? yes
VT_DYN_IGNORE_NODBG Disable instrumentation of functions which have no debug information? no
VT_DYN_INNER_LOOPS Instrument inner loops within outer loops?
(implies VT_DYN_OUTER_LOOPS=yes)		 no
VT_DYN_LOOP_ITERS Instrument loop iterations?
(implies VT_DYN_OUTER_LOOPS=yes)		 no
VT_DYN_OUTER_LOOPS Instrument outer loops within functions? no
VT_DYN_SHLIBS Colon-separated list of shared libraries for Dyninst instrumentation.
⇒ Section 2.6		 -

Filtering, Grouping
VT_FILTER_SPEC Name of function/region filter file.
⇒ Section 5.1		 -
VT_GROUPS_SPEC Name of function grouping file.
⇒ Section 5.3		 -
VT_JAVA_FILTER_SPEC Name of Java specific filter file.
⇒ Section 5.2		 -
VT_JAVA_GROUP_CLASSES Create a group for each Java class automatically? yes
VT_MAX_STACK_DEPTH Maximum number of stack level to be traced.
(0 = unlimited)		 0
VT_ONOFF_CHECK_STACK_BALANCE Check stack level balance when switching tracing on/off.
⇒ Section 2.4.2		 yes

Symbol List
VT_GNU_NM Command to list symbols from object files.
⇒ Section 2.3		 nm
VT_GNU_NMFILE Name of file with symbol list information.
⇒ Section 2.3		 -

The variables VT_PFORM_GDIR, VT_PFORM_LDIR, VT_FILE_PREFIX may contain (sub)strings of the form $XYZ or ${XYZ} where XYZ is the name of another environment variable. Evaluation of the environment variable is done at measurement runtime.

When you use these environment variables, make sure that they have the same value for all processes of your application on all nodes of your cluster. Some cluster environments do not automatically transfer your environment when executing parts of your job on remote nodes of the cluster, and you may need to explicitly set and export them in batch job submission scripts.


Influencing Trace Buffer Size

The default values of the environment variables VT_BUFFER_SIZE and VT_MAX_FLUSHES limit the internal buffer of VampirTrace to 32 MB per process and the number of times that the buffer is flushed to 1, respectively. Events that are to be recorded after the limit has been reached are no longer written into the trace file. The environment variables apply to every process of a parallel application, meaning that applications with n processes will typically create trace files n times the size of a serial application.

To remove the limit and get a complete trace of an application, set VT_MAX_FLUSHES to 0. This causes VampirTrace to always write the buffer to disk when it is full. To change the size of the buffer, use the environment variable VT_BUFFER_SIZE. The optimal value for this variable depends on the application which is to be traced. Setting a small value will increase the memory available to the application, but will trigger frequent buffer flushes by VampirTrace. These buffer flushes can significantly change the behavior of the application. On the other hand, setting a large value, like 2G, will minimize buffer flushes by VampirTrace, but decrease the memory available to the application. If not enough memory is available to hold the VampirTrace buffer and the application data, parts of the application may be swapped to disk, leading to a significant change in the behavior of the application.

In multi-threaded applications a single buffer cannot be shared across a process and the associated threads for performance reasons. Thus independent buffers are created for every process and thread, at which the process buffer size is 70% and the thread buffer size is 10% of the value set in VT_BUFFER_SIZE. The buffer size of processes and threads can be explicitly specified setting the environment variable VT_THREAD_BUFFER_SIZE, which defines the buffer size of a thread, whereas the buffer size of a process is then defined by the value of VT_BUFFER_SIZE. The total memory consumption of the application is calculated as follows (assuming that every process has the same number of threads):

a)
M = N * VT_BUFFER_SIZE * 0.7 + N * T * VT_BUFFER_SIZE * 0.1
(VT_THREAD_BUFFER_SIZE is not specified)
b)
M = N * VT_BUFFER_SIZE + N * T * VT_THREAD_BUFFER_SIZE
(VT_THREAD_BUFFER_SIZE is specified)
M ... total allocated memory   N ... number of processes   T ... number of threads per process


Note that you can decrease the size of trace files significantly by using the runtime function filtering as explained in Section 5.1.


Profiling an Application

Profiling an application collects aggregated information about certain events during a program run, whereas tracing records information about individual events. Profiling can therefore be used to get a summary of the program activity and to detect events that are called very often. The profiling information can also be used to generate filter rules to reduce the trace file size (⇒ Section 5.1).

To profile an application set the variable VT_MODE to STAT. Setting VT_MODE to STAT:TRACE tells VampirTrace to perform tracing and profiling at the same time. By setting the variable VT_STAT_PROPS the user can influence whether functions, messages, and/or collective operations shall be profiled. See Section 3.2 for information about these environment variables.


Unification of Local Traces

After a run of an instrumented application the traces of the single processes need to be unified in terms of timestamps and event IDs. In most cases, this happens automatically. If the environment variable VT_UNIFY is set to no or under certain circumstances it is necessary to perform unification of local traces manually. To do this, use the following command:



% vtunify <prefix>


If VampirTrace was built with support for OpenMP and/or MPI, it is possible to speedup the unification of local traces significantly. To distribute the unification on multible processes the MPI parallel version vtunify-mpi can be used as follow:



% mpirun -np <nranks> vtunify-mpi <prefix>


Furthermore, both tools vtunify and vtunify-mpi are capable to open additional OpenMP threads for unification. The number of threads can be specified by the OMP_NUM_THREADS environment variable.


Synchronized Buffer Flush

When tracing an application, VampirTrace temporarily stores the recorded events in a trace buffer. Typically, if a buffer of a process or thread has reached its maximum fill level, the buffer has to be flushed and other processes or threads maybe have to wait for this process or thread. This will result in an asynchronous runtime behavior.
To avoid this problem, VampirTrace provides a buffer flush in a synchronized manner. That means, if one buffer has reached its minimum buffer fill level VT_SYNC_FLUSH_LEVEL (⇒ Section 3.2), all buffers will be flushed. This buffer flush is only available at appropriate points in the program flow. Currently, VampirTrace makes use of all MPI collective functions associated with MPI_COMM_WORLD. Use the environment variable VT_SYNC_FLUSH to enable synchronized buffer flush.


Enhanced Timer Synchronization

Especially on cluster environments, where each process has its own local timer, tracing relies on precisely synchronized timers. Therefore, VampirTrace provides several mechanisms for timer synchronization. The default synchronization scheme is a linear synchronization at the very begin and the very end of a trace run with a master-slave communication pattern.
However, this way of synchronization can become to imprecise for long trace runs. Therefore, we recommend the usage of the enhanced timer synchronization scheme of VampirTrace. This scheme inserts additional synchronization phases at appropriate points in the program flow. Currently, VampirTrace makes use of all MPI collective functions associated with MPI_COMM_WORLD.
To enable this synchronization scheme, a LAPACK library with C wrapper support has to be provided for VampirTrace and the environment variable VT_ETIMESYNC (⇒ Section 3.2) has to be set before the tracing.
The length of the interval between two successive synchronization phases can be adjusted with VT_ETIMESYNC_INTV.
The following LAPACK libraries provide a C-LAPACK API that can be used by VampirTrace for the enhanced timer synchronization:

Note:

Systems equipped with a global timer do not need timer synchronization.

Note:

It is recommended to combine enhanced timer synchronization and synchronized buffer flush.

Note:

Be aware that the asynchronous behavior of the application will be disturbed since VampirTrace makes use of asynchronous MPI collective functions for timer synchronization and synchronized buffer flush.
Only make use of these approaches, if your application does not rely on an asynchronous behavior! Otherwise, keep this fact in mind during the process of performance analysis.


Environment Configuration Using VTSetup

In order to ease the process of configuring the runtime environment, the graphical tool vtsetup has been added to the VampirTrace toolset. With the help of a graphical user interface, required environment variables can be configured. The following option categories can be managed:

Furthermore, the user is granted more fine-grained control by activating the Advanced View button. The configuration can be saved to an XML file. After successfull configuration, the application can be launched directly or a script can be generated for manual execution.

Recording Additional Events and Counters


Hardware Performance Counters

If VampirTrace has been built with hardware counter support , it is capable of recording hardware counter information as part of the event records. To request the measurement of certain counters, the user is required to set the environment variable VT_METRICS. The variable should contain a colon-separated list of counter names or a predefined platform-specific group.

The user can leave the environment variable unset to indicate that no counters are requested. If any of the requested counters are not recognized or the full list of counters cannot be recorded due to hardware resource limits, program execution will be aborted with an error message.

PAPI Hardware Performance Counters

If the PAPI library is used to access hardware performance counters, metric names can be any PAPI preset names or PAPI native counter names. For example, set



VT_METRICS=PAPI_FP_OPS:PAPI_L2_TCM:!CPU_TEMP1


to record the number of floating point instructions and level 2 cache misses (PAPI preset counters), cpu temperature from the lm_sensors component. The leading exclamation mark let CPU_TEMP1 be interpreted as absolute value counter. See Section C.1 for a full list of PAPI preset counters.

CPC Hardware Performance Counters

On Sun Solaris operating systems VampirTrace can make use of the CPC performance counter library to query the processor's hardware performance counters. The counters which are actually available on your platform can be queried with the tool vtcpcavail. The listed names can then be used within VT_METRICS to tell VampirTrace which counters to record.

NEC SX Hardware Performance Counters

On NEC SX machines VampirTrace uses special register calls to query the processor's hardware counters. Use VT_METRICS to specify the counters that have to be recorded. See Section C.3 for a full list of NEC SX hardware performance counters.


Resource Usage Counters

The Unix system call getrusage provides information about consumed resources and operating system events of processes such as user/system time, received signals, and context switches.

If VampirTrace has been built with resource usage support, it is able to record this information as performance counters to the trace. You can enable tracing of specific resource counters by setting the environment variable VT_RUSAGE to a colon-separated list of counter names, as specified in Section C.4. For example, set



VT_RUSAGE=ru_stime:ru_majflt


to record the system time consumed by each process and the number of page faults. Alternatively, one can set this variable to the value all to enable recording of all 16 resource usage counters. Note that not all counters are supported by all Unix operating systems. Linux 2.6 kernels, for example, support only resource information for six of them. See Section C.4 and the manual page of getrusage for details.

The resource usage counters are not recorded at every event. They are only read if 100ms have passed since the last sampling. The interval can be changed by setting VT_RUSAGE_INTV to the number of desired milliseconds. Setting VT_RUSAGE_INTV to zero leads to sampling resource usage counters at every event, which may introduce a large runtime overhead. Note that in most cases the operating system does not update the resource usage information at the same high frequency as the hardware performance counters. Setting VT_RUSAGE_INTV to a value less than 10ms does usually not improve the granularity.

Be aware that, when using the resource usage counters for multi-threaded programs, the information displayed is valid for the whole process and not for each single thread.


Memory Allocation Counter

Calls to functions which reside in external libraries can be intercepted by implementing identical functions and linking them before the external library. Such ``wrapper functions'' can record the parameters and return values of the library functions.

If VampirTrace has been built with memory allocation tracing support , it uses this technique for recording calls to memory (de)allocation functions of the standard C library, which are executed by the application. The following functions are intercepted by VampirTrace:

malloc memalign
calloc posix_memalign
realloc valloc
free  

The gathered information will be saved as counter which indicates the current memory allocated in bytes. To request the measurement of the application's allocated memory, the user must set the environment variable VT_MEMTRACE to yes.

Note:

VampirTrace currently does not support memory allocation tracing for threaded programs, e.g., programs parallelized with OpenMP or Pthreads!


CPU ID Counter

The GNU LIBC implementation provides a function to determine the core id of a CPU on which the calling thread is running. VampirTrace uses this functionality to record the current core identifier as counter. This feature can be activated by setting the environment variable VT_CPUIDTRACE to yes.

Note:

To use this feature you need the GNU LIBC implementation at least in version 2.6.


NVIDIA CUDA

When tracing CUDA applications, only user events and functions are recorded, which are automatically or manually instrumented. CUDA API functions will not be traced by default. To enable tracing of CUDA runtime and driver API functions and CUDA device activities (like kernel execution and memory copies) build VampirTrace with CUDA support and set the following environment variable:

export VT_GPUTRACE=[yes|default|no]

To enable a particular composition of CUDA measurement features the variable should contain a comma-separated list of available CUDA measurement options.

export VT_GPUTRACE=option1,option2,option2,...

cuda enable CUDA (needed to use CUDA runtime API wrapper)
  (OpenCL is available in VampirTrace GPU beta releases)
cupti use the CUPTI interface instead of the library wrapper
runtime CUDA runtime API
driver CUDA driver API
kernel CUDA kernels
concurrent force recording of concurrent kernels with CUPTI
idle GPU compute idle time
pure_idle GPU idle time - considering data transfers as not idle
  (acts as idle for CUDA runtime API wrapper)
memcpy CUDA memory copies
sync enable recording of synchronization for tracing via CUPTI
stream_reuse force reusing of CUDA streams after cudaStreamDestroy()
memusage CUDA memory allocation
debug CUDA tracing debug mode
error CUDA errors will exit the program
yes|default same as ``cuda,runtime,kernel,memcpy''
no disable CUDA measurement

Since CUDA Toolkit 4.1 the CUDA Profiling and Tool Interface (CUPTI) allows capturing of CUDA device activities. VampirTrace trace has currently two methods to trace the CUDA runtime API and corresponding GPU activities: traditional library wrapping with CUDA events for GPU activity measurement and tracing via the CUPTI interface. Several features are just implemented in the library wrapping approach, whereas the CUPTI measurement brings new possibilities and occasionally more accuracy.


The new environment variable VT_GPUTRACE replaces several previously available environment variables. However, there are still additional feature switches implemented as environment variables to further refine CUDA tracing (the default is bold):

VT_GPUTRACE_KERNEL=[yes|2]
Tracing of CUDA kernels can be enabled with 'yes'. This is the same as adding the option kernel to VT_GPUTRACE. With '2' additional kernel counters are captured. (CUPTI tracing only)

VT_GPUTRACE_SYNC=[0|1|2|3]
Controls how VampirTrace handles synchronizing CUDA API calls, especially CUDA memory copies and CUDA device synchronization. At level 0 only the CUDA calls will be executed, messages will be displayed from the beginning to the end of the cudaMemcpy, regardless how long the cudaMemcpy call has to wait for a kernel until the actual data transfer starts. At level 1 the cudaMemcpy will be split into an additional synchronization and the actual data transfer in order to monitor the data transfer correctly. The additional synchronization does not affect the program execution significantly and will not be shown in the trace. At level 2 the additional synchronization will be exposed to the user. This allows a better view on the application execution, showing how much time is actually spent waiting for the GPU to complete. Level 3 will further use the synchronization to flush the internal task buffer and perform a timer synchronization between GPU and host. This introduces a minimal overhead but increases timer precision and prevents flushes elsewhere in the trace.

VT_CUPTI_METRICS
Capture CUDA CUPTI counters. Metrics are separated by default with '':`` or user specified by VT_METRICS_SEP.
Example: VT_CUPTI_METRICS=local_store:local_load

VT_CUPTI_EVENTS_SAMPLING=[yes|no]
Poll for CUPTI counter values during kernel execution, if set to yes.

VT_GPUTRACE_MEMUSAGE=[yes|2]
Record GPU memory usage as counter ``gpu_mem_usage``, if set to yes, which is the same as adding the option memusage to VT_GPUTRACE. With '2' missing cudaFree() calls are printed to stderr.

Every CUDA stream, which is executed on a cuda-capable device and used during program execution, creates an own thread. ``CUDA-Threads'' can contain CUDA communication, kernel and counter records and have the following notation:

CUDA[device:stream] process:thread

Due to an issue with CUPTI, the device is not always properly shown (device is displayed as ?). The CUDA stream number is increasing, beginning with the default stream 1. The stream number provided by CUPTI might not be evenly increasing. CUDA streams without records will not be written.

If CUDA libraries are used, which create CUDA streams themselves, many CUDA threads (CUDA streams per device) can appear in a program trace. In that case, it may be useful to force reusing of CUDA streams (add option stream_reuse to VT_GPUTRACE). This enables VampirTrace to reuse an existing thread buffer of a destroyed CUDA stream and therewith reduces the number of separate CUDA threads (or streams) in the trace. The CUDA stream number will then be missing in the CUDA thread notation.

As kernels and asynchronous memory copies are executed asynchronously on the CUDA device, information about these activities will be buffered until a synchronizing CUDA API function call or the program exits. Every used CUDA device and its corresponding host thread has an own buffer (8192 bytes by default), when CUDA tracing is done via the CUDA runtime API wrapper. If the buffer is full, it will be flushed immediately. When using CUDA tracing via CUPTI every CUDA context creation initiates the allocation of an own buffer (65536 bytes by default). If the buffer is full, further records will be dropped and a warning will be shown in stderr output. The buffer size can be specified in bytes with the environment variable VT_CUDATRACE_BUFFER_SIZE.

Several new region groups have been introduced:

CUDART_API CUDA runtime API calls
CUDRV_API CUDA driver API calls
CUDA_SYNC CUDA synchronization
CUDA_KERNEL CUDA kernels (device functions) can only appear on ``CUDA-Threads''
GPU_IDLE GPU compute idle time - the CUDA device does not run any kernel currently (shown in first used stream of the device)
VT_CUDA Measurement overhead (write CUDA events, check current device, etc.)

Tracing CUDA Runtime API via CUPTI

If the VampirTrace CUDA runtime API wrapper and CUPTI are configured during the VampirTrace build process, the option cupti has to be added to VT_GPUTRACE to enable CUDA runtime API tracing via CUPTI. In that case the CUDA runtime library should be preloaded to reduce tracing overhead (the dynamic linker can use LD_PRELOAD=libcudart.so). Otherwise the library wrapper intercepts every CUDA runtime API call and makes a short but unnecessary check, whether it is enabled.

Synchronous CUDA peer-to-peer memory copies will only be recorded, if the sync option is set and the synchronization level is 3 (default).

CUDA Runtime API Wrapper Particularities

CUDA tracing via this method will always record the CUDA runtime API calls. It is not possible to only record kernels, memory copies or memory usage. CUDA driver API programs cannot be traced with the CUDA runtime API wrapper.

Until CUDA Toolkit 4.2 the usage of CUDA events between asynchronous tasks serializes their on-device execution. As VampirTrace uses CUDA events for time measurement and asynchronous tasks may overlap (depends on the CUDA device capability), there might be a sensible impact on the program flow. CUDA 5 removes this restriction.

Counter via CUDA API

If VT_GPUTRACE_MEMUSAGE is enabled, CUDA memory allocations on the GPU will be tracked to write the GPU memory usage counter gpu_mem_usage. The counter values will be written directly to the default CUDA stream '1'. This stream will be created, if it does not exist and does not have to contain any other CUDA device activities. If the environment variable is set to 2, missing cudaFree() calls will be printed to stderr.

With kernel tracing enabled there are three counters, which provide information about the kernel's grid, block and thread composition: blocks_per_grid, threads_per_block, threads_per_kernel. With CUPTI tracing additional kernel counters are available: static and dynamic shared memory, total local memory and registers per thread (VT_GPUTRACE_KERNEL=2).

CUDA Performance Counters via CUPTI Events

To capture performance counters in CUDA applications, CUPTI events can be specified with the environment variable VT_CUPTI_METRICS. Counters are separated by default with '':`` or user specified by VT_METRICS_SEP. The CUPTI User's Guide - Event Reference provides information about the available counters. Alternatively set VT_CUPTI_METRICS=help to show a list of available counters (help_long to print the counter description as well). This will only take effect, when a kernel is about to be executed.

Compile and Link CUDA Applications

Use the VampirTrace compiler wrapper vtnvcc instead of nvcc to compile the CUDA application, which does automatic source code instrumentation.


GCC4.3 and OpenMP:
Use the flags -vt:opari -nodecl -Xcompiler=-fopenmp with vtnvcc to compile the OpenMP CUDA application.


CUDA 3.1:
The CUDA runtime library 3.1 creates a conflict with zlib. A workaround is to replace all gcc/g++ calls with the VampirTrace compiler wrappers (vtcc/vtc++) and pass the following additional flags to nvcc for compilation of the kernels: 

  -I$VT_INSTALL_PATH/include/vampirtrace
  -L$VT_INSTALL_PATH/lib 
  -Xcompiler=-g,-finstrument-functions,-pthread
  -lvt -lopen-trace-format -lcudart -lz -ldl -lm
$VT_INSTALL_PATH is the path to the VampirTrace installation directory. It is not necessary to specify the VampirTrace include and library path, if it is installed in the default directory.

This uses automatic compiler instrumentation (-finstrument-functions) and the standard VampirTrace library. Replace the -lvt with -lvt-mt for multithreaded, -lvt-mpi for MPI and -lvt-hyb for multithreaded MPI applications. In this case the CUDA runtime library is linked before the zlib.

If the application is linked with gcc/g++, the linking command has to ensure, that the respective VampirTrace library is linked before the CUDA runtime library libcudart.so (check e.g. with ``ldd executable''). Using the VampirTrace compiler wrappers (vtcc/vtc++) for linking is the easiest way to ensure correct linking of the VampirTrace library.

With the library tracing mechanism described in section 2.9, it is possible to trace CUDA applications without recompiling or relinking. There are only events written for Runtime API calls, kernels and communication between host and device.

Tracing the NVIDIA CUDA Sample Applications

CUDA 3.x and 4.x:
To get some example traces, replace the compiler commands in the common Makefile include file (common/common.mk) with the corresponding VampirTrace compiler wrappers (⇒2.1) for automatic instrumentation: 
  # Compilers
  NVCC := vtnvcc
  CXX  := vtc++
  CC   := vtcc
  LINK := vtc++ #-vt:mt

CUDA 5.0:
Set the following environment variables for automatic instrumentation before running make: 

  export GCC=vtc++ #-vt:mt
  export NVCC=vtnvcc #-vt:mt

Use the compiler switches for MPI, multi-threaded and hybrid programs, if necessary (e.g. the CUDA SDK example simpleMultiGPU is a multi-threaded program, which needs to be linked with a multi-threaded VampirTrace library).

Recording Concurrent Kernels (CUDA 5)

Since CUDA 5 it is possible to record concurrently executed kernels on the GPU. The VampirTrace CUDA runtime API wrapper uses CUDA events for GPU activity time measurement and is therefore by default enabled for recording concurrent kernels. The NVIDIA CUPTI library provides two possibilities for measuring kernels. If a CUDA application creates the second CUDA stream, the activity buffer will be flushed, the light-weight kernel recording disabled and concurrent kernel recording enabled. To force concurrent kernel support at VampirTrace CUDA initialization add the GPU tracing option concurrent.



Notes:
For 32-bit systems VampirTrace has to be configured with the 32-bit version of the CUDA runtime library. If the link test fails, use the following configure option : 

  --with-cuda-lib-dir=$CUDA_INSTALL_PATH/lib

To build CUPTI support on 32-bit systems (or for CUPTI 1.0), VampirTrace has to be configured with the 32-bit version of the CUPTI library. If the link test fails, use the following configure option : 

  --with-cupti-lib-dir=$CUPTI_INSTALL_PATH/lib

VampirTrace CUDA support has been successfully tested with CUDA toolkit version 3.x, 4.x and 5.0.


Pthread API Calls

When tracing applications with Pthreads, only user events and functions are recorded which are automatically or manually instrumented. Pthread API functions will not be traced by default.
To enable tracing of all C-Pthread API functions include the header vt_user.h and compile the instrumented sources with -DVTRACE_PTHREAD. 
C/C++:
           #include "vt_user.h"


% vtcc -DVTRACE_PTHREAD hello.c -o hello


Note:

Currently, Pthread instrumentation is only available for C/C++.


Plugin Counter Metrics

Plugin Counter add additional metrics to VampirTrace. They highly depend on the plugins, which are installed on your system. Every plugin should provide a README, which should be checked for available metrics. Once you have downloaded and compiled a plugin, copy the resulting library to a folder, which is part of your LD_LIBRARY_PATH. To enable the tracing of a specific metric, you should set the environment variable VT_PLUGIN_CNTR_METRICS. It is set in the following manner 
export VT_PLUGIN_CNTR_METRICS=<library_name>_<event_name>
If you have for example a library named libKswEvents.so with the event page_faults, the you can set it with 
export VT_PLUGIN_CNTR_METRICS=KswEvents_page_faults
Visit http://www.tu-dresden.de/zih/vampirtrace/plugin_counter for documentation and examples.

Note:

Multiple events can be concatenated by using colons.


I/O Calls

If VampirTrace has been built with I/O tracing support , it uses the same technique as used to intercept memory (de)allocation functions (⇒ Section 4.3) for recording calls to I/O functions of the standard C library, which are executed by the application. The following functions are intercepted by VampirTrace:

close creat creat64 dup
dup2 fclose fcntl fdopen
fgetc fgets flockfile fopen
fopen64 fprintf fputc fputs
fread fscanf fseek fseeko
fseeko64 fsetpos fsetpos64 ftrylockfile
funlockfile fwrite getc gets
lockf lseek lseek64 open
open64 pread pread64 putc
puts pwrite pwrite64 read
readv rewind unlink write
writev      

The gathered information will be saved as I/O event records in the trace file. This feature has to be activated for each tracing run by setting the environment variable VT_IOTRACE to yes.

If you'd like to experiment with some other I/O library, set the environment variable VT_IOLIB_PATHNAME to the alternative one. Beware that this library must provide all I/O functions mentioned above otherwise VampirTrace will abort. Setting the environment variable VT_IOTRACE_EXTENDED to yes enables the collection of additional function arguments for some of the I/O function mentioned above. For example, this option stores offsets for pwrite and pread additionally to the I/O event record. Enabling VT_IOTRACE_EXTENDED automatically enables VT_IOTRACE.


Child Process Execution Calls

In addition to the memory allocation tracing (⇒ Section 4.3) and I/O tracing (⇒ Section 4.8), VampirTrace uses the library wrapping technique also to intercept functions of the standard C library for creating and controling child processes. These functions are:

execl execvp fork waitid
execlp execve system wait3
execle execvpe wait wait4
execv fexecve waitpid  

When VampirTrace detects a call of an exec function, the current trace file is closed before executing the new program. If the executed program is also instrumented with VampirTrace, it will create a different trace file. Note that VampirTrace aborts if the exec function returns unsuccessfully. Calling fork in an instrumented program creates an additional process in the same trace file. Using this feature requires building VampirTrace with support for tracing LIBC functions for creating and , and setting the environment variable VT_EXECTRACE to yes.


MPI Correctness Checking Using UniMCI

VampirTrace supports the recording of MPI correctness events, e.g., usage of invalid MPI requests. This is implemented by using the Universal MPI Correctness Interface (UniMCI), which provides an interface between tools like VampirTrace and existing runtime MPI correctness checking tools. Correctness events are stored as markers in the trace file and are visualized by Vampir.

If VampirTrace is built with UniMCI support, the user only has to enable MPI correctness checking. This is done by merely setting the environment variable VT_MPICHECK to yes. Further, if your application crashes due to an MPI error you should set VT_MPICHECK_ERREXIT to yes. This environmental variable forces VampirTrace to write its trace to disk and exit afterwards. As a result, the trace with the detected error is stored before the application might crash.

To install VampirTrace with correctness checking support it is necessary to have UniMCI installed on your system. UniMCI in turn requires you to have a supported MPI correctness checking tool installed, currently only the tool Marmot is known to have UniMCI support. So all in all you should use the following order to install with correctness checking support:

  1. Marmot
    (see http://www.hlrs.de/organization/av/amt/research/marmot)
  2. UniMCI
    (see http://www.tu-dresden.de/zih/unimci)
  3. VampirTrace
    (see http://www.tu-dresden.de/zih/vampirtrace)

Information on how to install Marmot and UniMCI is given in their respective manuals. VampirTrace will automatically detect an UniMCI installation if the unimci-config tool is in path.


User-defined Counters

In addition to the manual instrumentation (⇒ Section 2.4), the VampirTrace API provides instrumentation calls which allow recording of program variable values (e.g. iteration counts, calculation results, ...) or any other numerical quantity. A user-defined counter is identified by its name, the counter group it belongs to, the type of its value (integer or floating-point) and the unit that the value is quoted (e.g. ``GFlop/sec'').

The VT_COUNT_GROUP_DEF and VT_COUNT_DEF instrumentation calls can be used to define counter groups and counters: 

Fortran:
           #include "vt_user.inc"
           integer :: id, gid
           VT_COUNT_GROUP_DEF('name', gid)
           VT_COUNT_DEF('name', 'unit', type, gid, id)


C/C++:
           #include "vt_user.h"
           unsigned int id, gid;
           gid = VT_COUNT_GROUP_DEF("name");
           id = VT_COUNT_DEF("name", "unit", type, gid);

The definition of a counter group is optional. If no special counter group is desired, the default group ``User'' can be used. In this case, set the parameter gid of VT_COUNT_DEF() to VT_COUNT_DEFGROUP.

The third parameter type of VT_COUNT_DEF specifies the data type of the counter value. To record a value for any of the defined counters the corresponding instrumentation call VT_COUNT_*_VAL must be invoked.

Fortran:    
Type Count call Data type
VT_COUNT_TYPE_INTEGER VT_COUNT_INTEGER_VAL integer (4 byte)
VT_COUNT_TYPE_INTEGER8 VT_COUNT_INTEGER8_VAL integer (8 byte)
VT_COUNT_TYPE_REAL VT_COUNT_REAL_VAL real
VT_COUNT_TYPE_DOUBLE VT_COUNT_DOUBLE_VAL double precision

C/C++:    
Type Count call Data type
VT_COUNT_TYPE_SIGNED VT_COUNT_SIGNED_VAL signed int (max. 64-bit)
VT_COUNT_TYPE_UNSIGNED VT_COUNT_UNSIGNED_VAL unsigned int (max. 64-bit)
VT_COUNT_TYPE_FLOAT VT_COUNT_FLOAT_VAL float
VT_COUNT_TYPE_DOUBLE VT_COUNT_DOUBLE_VAL double

The following example records the loop index i: 

Fortran:

  #include "vt_user.inc"

  program main
  integer :: i, cid, cgid

  VT_COUNT_GROUP_DEF('loopindex', cgid)
  VT_COUNT_DEF('i', '#', VT_COUNT_TYPE_INTEGER, cgid, cid)

  do i=1,100
    VT_COUNT_INTEGER_VAL(cid, i)
  end do

  end program main


C/C++:

  #include "vt_user.h"

  int main() {
    unsigned int i, cid, cgid;

    cgid = VT_COUNT_GROUP_DEF('loopindex');
    cid = VT_COUNT_DEF("i", "#", VT_COUNT_TYPE_UNSIGNED,
                       cgid);

    for( i = 1; i <= 100; i++ ) {
      VT_COUNT_UNSIGNED_VAL(cid, i);
    }

    return 0;
  }

For all three languages the instrumented sources have to be compiled with -DVTRACE. Otherwise the VT_* calls are ignored.

Optionally, if the sources contain further VampirTrace API calls and only the calls for user-defined counters shall be disabled, then the sources have to be compiled with -DVTRACE_NO_COUNT in addition to -DVTRACE.


User-defined Markers

In addition to the manual instrumentation (⇒ Section 2.4), the VampirTrace API provides instrumentation calls which allow recording of special user information, which can be used to better identify parts of interest. A user-defined marker is identified by its name and type. 

Fortran:
           #include "vt_user.inc"
           integer :: mid
           VT_MARKER_DEF('name', type, mid)
           VT_MARKER(mid, 'text')

C/C++:
           #include "vt_user.h"
           unsigned int mid;
           mid = VT_MARKER_DEF("name",type);
           VT_MARKER(mid, "text");

Types for Fortran/C/C++:
           VT_MARKER_TYPE_ERROR
           VT_MARKER_TYPE_WARNING
           VT_MARKER_TYPE_HINT

For all three languages the instrumented sources have to be compiled with -DVTRACE. Otherwise the VT_* calls are ignored.

Optionally, if the sources contain further VampirTrace API calls and only the calls for user-defined markers shall be disabled, then the sources have to be compiled with -DVTRACE_NO_MARKER in addition to -DVTRACE.


User-defined Communcation

In addition to the manual instrumentation (⇒ Section 2.4), the VampirTrace API provides instrumentation calls which allow recording of special user information, which can be used to better identify parts of interest. A user-defined communication operation is defined by a communicator and a tag. The default communicator is VT_COMM_WORLD. Additionally, a user-defined communicator can be created using VT_COMM_DEF: 

Fortran:
         #include "vt_user.inc"
         integer :: cid
         VT_COMM_DEF('name', cid)
			
C/C++:
         #include "vt_user.h"
         unsigned cid;
         cid = VT_COMM_DEF("name", cid);

Using VT_SEND and VT_RECV the user can insert send and receive events into the trace: 

C/C++:
         int rank, size;
         MPI_Comm_rank(MPI_COMM_WORLD, &rank);
         MPI_Comm_size(MPI_COMM_WORLD, &size);

         if( rank == 0 )
         {
             for ( int i = 1; i < size; i++ )
             {
                 VT_SEND(VT_COMM_WORLD,i,100);
             }  
         }else
         {
             VT_RECV(VT_COMM_WORLD,rank,100);
         }

The calls are similar for Fortran.

As can be seen, the arguments to VT_SEND and VT_RECV are a communicator, a tag and the size of the message. The tag is required in order to identify both ends of a user-defined communication. Therefore it has to be globally unique for a given communicator and cannot be reused within a single communicator. Messages with duplicated tags will not be visible in the final trace.

For all three languages the instrumented sources have to be compiled with -DVTRACE. Otherwise the VT_* calls are ignored. Optionally, if the sources contain further VampirTrace API calls and only the calls for user-defined markers shall be disabled, then the sources have to be compiled with -DVTRACE_NO_MSG in addition to -DVTRACE.


Filtering & Grouping


Function Filtering

By default, all calls of instrumented functions will be traced, so that the resulting trace files can easily become very large. In order to decrease the size of a trace, VampirTrace allows the specification of filter directives before running an instrumented application. The user can decide on how often an instrumented function(group) shall be recorded to a trace file. To use a filter, the environment variable VT_FILTER_SPEC needs to be defined. It should contain the path and name of a file with filter directives specified as follows:

<function> - <limit> [S:<[min-]max-stack-level>] [R]
or
<groups> - <limit> [S:<[min-]max-stack-level>] [R] G
or
<function-call-path> - <limit> C


functions, groups Semicolon-separated list of
  functions/groups.
  (can contain wildcards)
   
function-call-path Semicolon-separated list of
  functions in a call path.
  (MUST NOT contain wildcards)
   
limit call limit
  Stop recording of functions/groups when
  the specified call limit is reached.
  (0 = don't record functions/groups,
  -1 record unlimited)
   


S:<[min-]max-stack-level>  
  minimum/maximum call stack level
  Don't record functions/groups called
  beyond the specified stack level
  boundaries.
  (values must be > 0, only valid if call
  limit is != 0)
   
R Attribute for recursive filtering.
  Don't record callees of filtered
  function/group.
   
G Attribute for filtering function groups.
   
C Attribute for filtering function a call path.
  (implies recursive filtering R)
   

Example: 

  add;sub;mul;div -- 1000
  MATH            -- 500 G
  *               -- 3000000 S:5-10

These filter directives cause that the functions add, sub, mul, and div will be recorded at most 1000 times. All the functions of the group MATH at most 500 times. The remaining functions * will only be recorded when they are called between call stack level 5 and 10 but at most 3000000 times.

Besides creating filter files manually, you can also use the vtfilter tool to generate them automatically. This tool reads a provided trace and decides whether a function should be filtered or not, based on the evaluation of

Call Path Specific Filtering

The 'C' attribute indicates that the listed functions specify a call path - a specific sequence of function calls. Recording of the last function in the list will be stopped if the specified call limit is reached. The call path must begin with the root function, typically main, and MUST NOT contain wildcards.

Example: 

  main;foo;bar -- 0 C

This filter directive causes that the function bar called from foo which prior was called from main will never be recorded. Since call path filtering impies recursiveness (see attribute R) all callee functions of this call path will be excluded from recording as well.

Rank Specific Filtering

An experimental extension allows rank specific filtering. Use @ clauses to restrict all following filters to the given ranks. The rank selection must be given as a list of <from> - <to> pairs or single values. Note that all rank specific rules are only effective after MPI_Init because the ranks are unknown before. The optional argument - OFF disables the given ranks completely, regardless of following filter rules. 

  @ 35 - 42 -- OFF
  @ 4 - 10, 20 - 29, 34
  foo;bar -- 2000
  * -- 0

The example defines two limits for the ranks 4 - 10, 20 - 29, and 34. The first line disables the ranks 35 - 42 completely.

Attention:

The rank specific rules are activated later than usual at MPI_Init, because the ranks are not available earlier. The special MPI routines MPI_Init, MPI_Init_thread, and MPI_Initialized cannot be filtered in this way.


Java Specific Filtering

For Java tracing there are additional possibilities of filtering. Firstly, there is a default filter applied. The rules can be found in the filter file <vt-install>/etc/ vt-java-default-filter.spec. Secondly, user-defined filters can be applied additionally by setting VT_JAVA_FILTER_SPEC to a file containing the rules.

The syntax of the filter rules is as follows: 

 <method|thread> <include|exclude> <filter string[;fs]...>

Filtering can be done on thread names and method names, defined by the first parameter. The second parameter determines whether the matching item shall be included for tracing or excluded from it. Multiple filter strings on a line have to be separated by ; and may contain occurences of * for wildcard matching.

The user-supplied filter rules will be applied before the default filter and the first match counts so it is possible to include items that would be excluded by the default filter otherwise.


Function Grouping

VampirTrace allows assigning functions/regions to a group. Groups can, for instance, be highlighted by different colors in Vampir displays. The following standard groups are created by VampirTrace:

Group name Contained functions/regions
MPI MPI functions
OMP OpenMP API function calls
OMP_SYNC OpenMP barriers
OMP_PREG OpenMP parallel regions
Pthreads Pthread API function calls
LIBC-EXEC LIBC function calls for creating and controling child processes (⇒ Section 4.9)
LIBC-I/O LIBC functions (⇒ Section 4.8)
LIBC-MALLOC LIBC memory (de)allocation functions (⇒ Section 4.3)
Application remaining instrumented functions and source code regions

Additionally, you can create your own groups, e.g., to better distinguish different phases of an application. To use function/region grouping set the environment variable VT_GROUPS_SPEC to the path of a file which contains the group assignments specified as follows: 

 <group>=<functions>


group group name
functions semicolon-seperated list of functions
  (can contain wildcards)

Example: 

  MATH=add;sub;mul;div
  USER=app_*

These group assignments associate the functions add, sub, mul, and div with group ``MATH'', and all functions with the prefix app_ are associated with group ``USER''.


VampirTrace Installation

Basics

Building VampirTrace is typically a combination of running configure and make. Execute the following commands to install VampirTrace from the directory at the top of the tree: 

% ./configure --prefix=/where/to/install
[...lots of output...]
% make all install

If you need special access for installing, you can execute make all as a user with write permissions in the build tree and a separate make install as a user with write permissions to the install tree.

However, for more details, also read the following instructions. Sometimes it might be necessary to provide ./configure with options, e.g.,  specifications of paths or compilers.

VampirTrace comes with example programs written in C, C++, and Fortran. They can be used to test different instrumentation types of the VampirTrace installation. You can find them in the directory examples of the VampirTrace package.

Note that you should compile VampirTrace with the same compiler you use for the application to trace.


Configure Options

Compilers and Options

Some systems require unusual options for compiling or linking which the configure script does not know. Run ./configure -help for details on some of the pertinent environment variables.

You can pass initial values for configuration parameters to configure by setting variables in the command line or in the environment. Here is an example: 

% ./configure CC=c89 CFLAGS=-O2 LIBS=-lposix

Installation Names

By default, make install will install the package's files in /usr/local/bin, /usr/local/include, etc. You can specify an installation prefix other than /usr/local by giving configure the option -prefix=PATH.

Optional Features

This a summary of the most important optional features. For a full list of all available features run ./configure -help.

-enable-compinst=TYPE
 
enable support for compiler instrumentation, e.g. gnu,pgi,pgi9,sun
default: automatically by configure. Note: Use pgi9 for PGI compiler version 9.0 or higher.

-enable-dyninst
 
enable support for Dyninst instrumentation, default: enable if found by configure. Note: Requires Dyninst[*] version 6.1 or higher!

-enable-dyninst-attlib
 
build shared library which attaches Dyninst to the running application, default: enable if Dyninst found by configure and system supports shared libraries

-enable-tauinst
 
enable support for automatic source code instrumentation by using TAU, default: enable if found by configure. Note: Requires PDToolkit[*] or TAU[*]!

-enable-cpuidtrace
 
enable CPU ID tracing support, default: enable if found by configure

-enable-libtrace=LIST
 
enable library tracing support (gen,exec,io,malloc,cudart), default: automatically by configure

-enable-exectrace
 
enable support for tracing LIBC functions for creating and controling child processes (e.g. execl,fork,system,wait) via library wrapping, default: enable

-enable-iotrace
 
enable support for tracing LIBC I/O functions (e.g. fopen,fclose,fread,fwrite) via library wrapping, default: enable

-enable-memtrace
 
enable support for tracing LIBC functions for memory de/allocation (e.g. malloc,realloc,free) via library wrapping, default: enable

-enable-cudartwrap
 
enable support for tracing the CUDA runtime API via library wrapping, default: enable if no CUPTI present

-enable-rutrace
 
enable resource usage tracing support, default: enable if found by configure

-enable-metrics=TYPE
 
enable support for hardware performance counter (papi,cpc,necsx), default: automatically by configure

-enable-zlib
 
enable ZLIB trace compression support, default: enable if found by configure

-enable-mpi
 
enable MPI support, default: enable if MPI found by configure

-enable-fmpi-lib
 
build the MPI Fortran support library, in case your system does not have a MPI Fortran library. default: enable if no MPI Fortran library found by configure

-enable-fmpi-handle-convert
 
do convert MPI handles, default: enable if MPI conversion functions found by configure

-enable-mpi2-thread
 
enable MPI-2 Thread support, default: enable if found by configure

-enable-mpi2-1sided
 
enable MPI-2 One-Sided Communication support, default: enable if found by configure

-enable-mpi2-extcoll
 
enable MPI-2 Extended Collective Operation support, default: enable if found by configure

-enable-mpi2-io
 
enable MPI-2 I/O support, default: enable if found configure

-enable-mpicheck
 
enable support for Universal MPI Correctness Interface (UniMCI), default: enable if unimci-config found by configure

-enable-etimesync
 
enable enhanced timer synchronization support, default: enable if C-LAPACK found by configure

-enable-threads=LIST
 
enable support for threads (pthread, omp), default: automatically by configure

-enable-java
 
enable Java support, default: enable if JVMTI found by configure

-enable-cupti
 
enable support for tracing CUDA via CUPTI, default: enable if found by configure

Important Optional Packages

This a summary of the most important optional features. For a full list of all available features run ./configure -help.

-with-platform=PLATFORM
 
configure for given platform (altix,bgl,bgp,crayt3e,crayx1,crayxt,
ibm,linux,macos,necsx,origin,sicortex,sun,generic), default: automatically by configure

-with-bitmode=32|64
 
specify bit mode

-with-options=FILE
 
load options from FILE, default: configure searches for a config file in config/defaults based on given platform and bitmode

-with-local-tmp-dir=DIR
 
give the path for node-local temporary directory to store local traces to, default: /tmp

If you would like to use an external version of OTF library, set:

-with-extern-otf
 
use external OTF library, default: not set
-with-extern-otf-dir=OTFDIR
 
give the path for OTF, default: /usr

-with-otf-flags=FLAGS
 
pass FLAGS to the OTF distribution configuration (only for internal OTF version)

-with-otf-lib=OTFLIB
 
use given otf lib, default: -lopen-trace-format -lz

If the supplied OTF library was built without zlib support then OTFLIB will be set to -lopen-trace-format.

-with-dyninst-dir=DYNIDIR
 
give the path for DYNINST, default: /usr

-with-dyninst-inc-dir=DYNIINCDIR
 
give the path for Dyninst-include files, default: DYNIDIR/include

-with-dyninst-lib-dir=DYNILIBDIR
 
give the path for Dyninst-libraries, default: DYNIDIR/lib

-with-dyninst-lib=DYNILIB
 
use given Dyninst lib, default: -ldyninstAPI

-with-tau-instrumentor=TAUINSTUMENTOR
 
give the command for the TAU instrumentor, default: tau_instrumentor

-with-pdt-cparse=PDTCPARSE
 
give the command for PDT C source code parser, default: cparse

-with-pdt-cxxparse=PDTCXXPARSE
 
give the command for PDT C++ source code parser, default: cxxparse

-with-pdt-fparse=PDTFPARSE
 
give the command for PDT Fortran source code parser, default: f95parse, f90parse, or gfparse

-with-papi-dir=PAPIDIR
 
give the path for PAPI, default: /usr

-with-cpc-dir=CPCDIR
 
give the path for CPC, default: /usr

If you have not specified the environment variable MPICC (MPI compiler command) use the following options to set the location of your MPI installation:

-with-mpi-dir=MPIDIR
 
give the path for MPI, default: /usr/

-with-mpi-inc-dir=MPIINCDIR
 
give the path for MPI-include files,
default: MPIDIR/include/

-with-mpi-lib-dir=MPILIBDIR
 
give the path for MPI-libraries, default: MPIDIR/lib/

-with-mpi-lib
 
use given mpi lib

-with-pmpi-lib
 
use given pmpi lib

If your system does not have an MPI Fortran library set -enable-fmpi-lib (see above), otherwise set:

-with-fmpi-lib
 
use given fmpi lib

Use the following options to specify your MPI-implementation

-with-hpmpi
 
set MPI-libs for HP MPI

-with-pcmpi
 
set MPI-libs for Platform MPI

-with-intelmpi
 
set MPI-libs for Intel MPI

-with-intelmpi2
 
set MPI-libs for Intel MPI2

-with-lam
 
set MPI-libs for LAM/MPI

-with-mpibgl
 
set MPI-libs for IBM BG/L

-with-mpibgp
 
set MPI-libs for IBM BG/P

-with-mpich
 
set MPI-libs for MPICH

-with-mpich2
 
set MPI-libs for MPICH2

-with-mvapich
 
set MPI-libs for MVAPICH

-with-mvapich2
 
set MPI-libs for MVAPICH2

-with-mpisx
 
set MPI-libs for NEC MPI/SX

-with-mpisx-ew
 
set MPI-libs for NEC MPI/SX with 8 Byte Fortran Integer

-with-openmpi
 
set MPI-libs for Open MPI

-with-sgimpt
 
set MPI-libs for SGI MPT

-with-sunmpi
 
set MPI-libs for SUN MPI

-with-sunmpi-mt
 
set MPI-libs for SUN MPI-MT

To enable enhanced timer synchronization a LAPACK library with C wrapper support is needed:

-with-clapack-dir=LAPACKDIR
 
set the path for CLAPACK, default: /usr

-with-clapack-lib
 
set CLAPACK-libs, default: -lclapack -lcblas -lf2c

-with-clapack-acml
 
set CLAPACK-libs for ACML

-with-clapack-essl
 
set CLAPACK-libs for ESSL

-with-clapack-mkl
 
set CLAPACK-libs for MKL

-with-clapack-sunperf
 
set CLAPACK-libs for SUN Performance Library

To enable Java support the JVM Tool Interface (JVMTI) version 1.0 or higher is required:

-with-jvmti-dir=JVMTIDIR
 
give the path for JVMTI, default: $JAVA_HOME

-with-jvmti-inc-dir=JVMTIINCDIR
 
give the path for JVMTI-include files, default: JVMTI/include

To enable support for generating wrapper for 3th-Party libraries the C code parser CTool[*] is needed:

-with-ctool-dir=CTOOLDIR
 
give the path for CTool, default: /usr

-with-ctool-inc-dir=CTOOLINCDIR
 
give the path for CTool-include files, default: CTOOLDIR/include

-with-ctool-lib-dir=CTOOLLIBDIR
 
give the path for CTool-libraries, default: CTOOLDIR/lib

-with-ctool-lib=CTOOLLIB
 
use given CTool lib, default: automatically by configure

To enable support for CUDA API wrapping, the CUDA-Toolkit install path is needed:

-with-cuda-dir=CUDATKDIR
 
give the path for CUDA Toolkit, default: /usr/local/cuda
-with-cuda-inc-dir=CUDATKINCDIR
 
give the path for CUDA Toolkit-include files, default: CUDATKDIR/include
-with-cuda-lib-dir=CUDATKLIBDIR
 
give the path for CUDA Toolkit-libraries, default: CUDATKDIR/lib64
-with-cudart-lib=CUDARTLIB
 
use given cudart lib, default: -lcudart
-with-cudart-shlib=CUDARTSHLIB
 
give the pathname for the shared CUDA runtime library, default: automatically by configure

To enable support for CUPTI features, the CUPTI install path is needed:

-with-cupti-dir=CUPTIDIR
 
give the path for CUPTI, default: /usr
-with-cupti-inc-dir=CUPTIINCDIR
 
give the path for CUPTI-include files, default: CUPTIDIR/include
-with-cupti-lib-dir=CUPTILIBDIR
 
give the path for CUPTI-libraries, default: CUPTIDIR/lib64
-with-cupti-lib=CUPTILIB
 
use given cupti lib, default: -lcupti

Cross Compilation

Building VampirTrace on cross compilation platforms needs some special attention. The compiler wrappers, OPARI, and the Library Wrapper Generator are built for the front-end (build system) whereas the the VampirTrace libraries, vtdyn, vtunify, and vtfilter are built for the back-end (host system). Some configure options which are of interest for cross compilation are shown below:

Examples:

BlueGene/P and BlueGene/Q: 

% ./configure --host=powerpc64-ibm-linux-gnu

Cray XK6: 

% ./configure --host=x86_64-cray-linux-gnu
              CC_FOR_BUILD=craycc
              CXX_FOR_BUILD=crayc++

NEC SX6: 

% ./configure --host=sx6-nec-superux14.1

Environment Set-Up

Add the bin subdirectory of the installation directory to your $PATH environment variable. To use VampirTrace with Dyninst, you will also need to add the lib subdirectory to your LD_LIBRARY_PATH environment variable:


for csh and tcsh: 

> setenv PATH <vt-install>/bin:$PATH
> setenv LD_LIBRARY_PATH <vt-install>/lib:$LD_LIBRARY_PATH
for bash and sh: 
% export PATH=<vt-install>/bin:$PATH
% export LD_LIBRARY_PATH=<vt-install>/lib:$LD_LIBRARY_PATH

Notes for Developers

Build from SVN

If you have checked out a developer's copy of VampirTrace (i.e. checked out from CVS), you should first run: 

% ./bootstrap [--otf-package <package>]
              [--version <version>]
Note that GNU Autoconf ≥2.60 and GNU Automake ≥1.9.6 are required. You can download them from http://www.gnu.org/software/autoconfand http://www.gnu.org/software/automake.

Command Reference


Compiler Wrappers (vtcc,vtcxx,vtfort) 

vtcc,vtcxx,vtfort - compiler wrappers for C, C++, Fortran

Syntax: vt<cc|cxx|fc> [options] ...

options:
  -vt:help            Show this help message.
  -vt:version         Show VampirTrace version.
  -vt:<cc|cxx|fc> <cmd>
                      Set the underlying compiler command.

  -vt:inst <insttype> Set the instrumentation type.

   possible values:

    compinst          fully-automatic by compiler
    manual            manual by using VampirTrace's API
    dyninst           binary by using Dyninst (www.dyninst.org)
    tauinst           automatic source code instrumentation by
                      using PDT/TAU

  -vt:inst-exclude-file-list <file>[,file,...]
                      Set list of source files to be excluded
                      from the automatic instrumentation by the
                      compiler or PDT/TAU.
                      (file names can contain wildcards)

  -vt:inst-exclude-file <file>
                      Set pathname of file containing a list of
                      source files to be excluded from the
                      automatic instrumentation by the compiler
                      or PDT/TAU.
                      (file names can contain wildcards, one file
                       name per line)

   Note when using an exclusion list for automatic compiler
   instrumentation:
   If a source file from the exclusion list is involved in a
   compile step, the instrumentation is disabled for this step.

  -vt:opari <!args>   Set options for OPARI command. (see
                      share/vampirtrace/doc/opari/Readme.html)

  -vt:opari-rcfile <file>
                      Set pathname of the OPARI resource file.
                      (default: opari.rc)

  -vt:opari-table <file>
                      Set pathname of the OPARI runtime table file.
                      (default: opari.tab.c)

  -vt:opari-exclude-file-list <file>[,file,...]
                      Set list of source files to be excluded from
                      the instrumentation of OpenMP constructs by
                      OPARI.
                      (file names can contain wildcards)

  -vt:opari-exclude-file <file>
                      Set pathname of file containing a list of
                      source files to be excluded from the
                      instrumentation of OpenMP constructs by OPARI.
                      (file names can contain wildcards, one file name
                       per line)

  -vt:noopari         Disable instrumentation of OpenMP contructs
                      by OPARI.

  -vt:<seq|mpi|mt|hyb>
                      Enforce application's parallelization type.
                      It's only necessary if it could not be determined
                      automatically based on underlying compiler and flags.
                      seq = sequential
                      mpi = parallel (uses MPI)
                      mt = parallel (uses OpenMP/POSIX threads)
                      hyb = hybrid parallel (MPI + Threads)
                      (default: automatically)

  -vt:tau <!args>     Set options for the TAU instrumentor 
                      command.

  -vt:pdt <!args>     Set options for the PDT parse command.

  -vt:preprocess      Preprocess the source files before parsing
                      by OPARI and/or PDT.

  -vt:cpp <cmd>       Set C preprocessor command.

  -vt:cppflags <[!]flags>
                      Set/add flags for the C preprocessor.

  -vt:verbose         Enable verbose mode.

  -vt:keepfiles       Keep intermediate files.

  -vt:reusefiles      Reuse intermediate files, if exist.

  -vt:show[me]        Do not invoke the underlying compiler.
                      Instead, show the command line that would be
                      executed to compile and link the program.

  -vt:showme-compile  Do not invoke the underlying compiler.
                      Instead, show the compiler flags that would be
                      supplied to the compiler.

  -vt:showme-link     Do not invoke the underlying compiler.
                      Instead, show the linker flags that would be
                      supplied to the compiler.

  See the man page for your underlying compiler for other 
  options that can be passed through 'vt<cc|cxx|fc>'.

Environment variables:
  VT_INST             Equivalent to '-vt:inst'
  VT_CC               Equivalent to '-vt:cc '
  VT_CXX              Equivalent to '-vt:cxx '
  VT_FC               Equivalent to '-vt:fc'
  VT_CFLAGS           C compiler flags
  VT_CXXFLAGS         C++ compiler flags
  VT_FCFLAGS          Fortran compiler flags
  VT_LDFLAGS          Linker flags
  VT_LIBS             Libraries to pass to the linker

  The corresponding command line options overwrite the 
  environment variables setting.

Examples:
  automatically instrumentation by compiler:

     vtcc -vt:cc gcc -vt:inst compinst -c foo.c -o foo.o
     vtcc -vt:cc gcc -vt:inst compinst -c bar.c -o bar.o
     vtcc -vt:cc gcc -vt:inst compinst foo.o bar.o -o foo

  manually instrumentation by using VT's API:

     vtfort -vt:inst manual foobar.F90 -o foobar -DVTRACE

  IMPORTANT: Fortran source files instrumented by VT's API
             have to be preprocessed by CPP.


Local Trace Unifier (vtunify) 

vtunify[-mpi] - local trace unifier for VampirTrace.

Syntax: vtunify[-mpi] [options] <input trace prefix>

options:
  -h, --help          Show this help message.

  -V, --version       Show VampirTrace version.

  -o PREFIX           Prefix of output trace filename.

  -f FILE             Function profile output filename.
                      (default=PREFIX.prof.txt)

  -k, --keeplocal     Don't remove input trace files.

  -p, --progress      Show progress.

  -v, --verbose       Increase output verbosity.
                      (can be used more than once)

  -q, --quiet         Enable quiet mode.
                      (only emergency output)

  --iofsl-servers LIST
                      Enable IOFSL mode where LIST contains a comma-separated
                      list of IOFSL server addresses.

  --iofsl-mode MODE   IOFSL mode (MULTIFILE or MULTIFILE_SPLIT).
                      (default: MULTIFILE_SPLIT)

  --iofsl-asyncio     Use asynchronous I/O in IOFSL mode.

  --stats             Unify only summarized information (*.stats), no events

  --nocompress        Don't compress output trace files.

  --nosnapshots       Don't create snapshots.

  --maxsnapshots N    Maximum number of snapshots.
                      (default: 1024)

  --nomsgmatch        Don't match messages.

  --droprecvs         Drop message receive events, if msg. matching
                      is enabled.


Binary Instrumentor (vtdyn) 

vtdyn - binary instrumentor (Dyninst mutator) for VampirTrace.

Syntax: vtdyn [options] <executable> [arguments ...]

options:
  -h, --help          Show this help message.

  -V, --version       Show VampirTrace version.

  -v, --verbose       Increase output verbosity.
                      (can be used more than once)

  -q, --quiet         Enable quiet mode.
                      (only emergency output)

  -o, --output FILE   Rewrite instrumented executable to specified pathname.

  -f, --filter FILE   Pathname of input filter file.

  -s, --shlibs SHLIBS[,...]
                      Comma-separated list of shared libraries which shall
                      also be instrumented.

  --outer-loops       Do instrument outer loops within functions.

  --inner-loops       Do instrument inner loops within outer loops.
                      (implies --outer-loops)

  --loop-iters        Do instrument loop iterations.
                      (implies --outer-loops)

  --ignore-nodbg      Don't instrument functions which have no debug
                      information.


Trace Filter Tool (vtfilter) 

vtfilter[-mpi] - filter tool for VampirTrace.

Syntax: 
  Generate a filter file:
    vtfilter[-mpi] --gen [gen-options] <input trace file>

  Filter a trace using an already existing filter file:
    vtfilter[-mpi] [--filt] [filt-options]
      --filter=<input filter file> <input trace file>

options:
  --gen               Generate a filter file.
                      See 'gen-options' below for valid options.

  --filt              Filter a trace using an already existing
                      filter file. (default)
                      See 'filt-options' below for valid options.

  -h, --help          Show this help message.

  -V, --version       Show VampirTrace version.

  -p, --progress      Show progress.

  -v, --verbose       Increase output verbosity.
                      (can be used more than once)

gen-options:
  -o, --output=FILE   Pathname of output filter file.

  -r, --reduce=N      Reduce the trace size to N percent of the
                      original size. The program relies on the
                      fact that the major part of the trace are
                      function calls. The approximation of size
                      will get worse with a rising percentage of
                      communication and other non function
                      calling or performance counter records.                           

  -l, --limit=N       Limit the number of calls for filtered
                      function to N.
                      (default: 0)                                         

  -s, --stats         Prints out the desired and the expected
                      percentage of file size.                                     

  -e, --exclude=FUNC[;FUNC;...]
                      Exclude certain functions from filtering.
                      A function name may contain wildcards.   

  --exclude-file=FILE Pathname of file containing a list of
                      functions to be excluded from filtering.                             

  -i, --include=FUNC[;FUNC;...]
                      Force to include certain functions into
                      the filter. A function name may contain
                      wildcards.             

  --include-file=FILE Pathname of file containing a list of
                       functions to be included into the filter.                            

  --include-callees   Automatically include callees of included
                      functions as well into the filter.                           

filt-options:
  -o, --output=FILE   Pathname of output trace file.

  -f, --filter=FILE   Pathname of input filter file.

  -s, --max-streams=N Maximum number of output streams.
                      (default: 0)
            vtfilter: Set this to 0 to get the same number of
                      output streams as input streams.                                     
        vtfilter-mpi: Set this to 0 to get the same number of
                      output streams as MPI processes used, but
                      at least the number of input streams.

  --max-file-handles=N
                      Maximum number of files that are allowed
                      to be open simultaneously.
                      (default: 256)

  --nocompress        Don't compress output trace files.


Library Wrapper Generator (vtlibwrapgen) 

vtlibwrapgen - library wrapper generator for VampirTrace.

Syntax: 
  Generate a library wrapper source file:
    vtlibwrapgen [gen-options] <input header file> 
                 [input header file...]

  Build a wrapper library from a generated source file:
    vtlibwrapgen --build [build-options] 
                 <input lib. wrapper source file>

options:
  --gen              Generate a library wrapper source file. 
                     This is the default behavior. See 
                     'gen-options' below for valid options.

  --build            Build a wrapper library from a generated 
                     source file. See 'build-options' below 
                     for valid options.

  -h, --help         Show this help message.

  -V, --version      Show VampirTrace version.

  -q, --quiet        Enable quiet mode. 
                     (only emergency output)

  -v, --verbose      Increase output verbosity.
                     (can be used more than once)

gen-options:
  -o, --output=FILE  Pathname of output wrapper source file.
                     (default: wrap.c)                      

  -l, --shlib=SHLIB  Pathname of shared library that contains 
                     the actual library functions.
                     (can be used more then once)

  -f, --filter=FILE  Pathname of input filter file.

  -g, --group=NAME   Separate function group name for wrapped 
                     functions.

  -s, --sysheader=FILE
                     Header file to be included additionally.

  --nocpp            Don't use preprocessor.

  --keepcppfile      Don't remove preprocessed header files.

  --cpp=CPP          C preprocessor command
                     (default: gcc -E)     

  --cppflags=CPPFLAGS 
                     C preprocessor flags, e.g. 
                     -I<include dir>

  --cppdir=DIR       Change to this preprocessing directory.

environment variables:
  VT_CPP             C preprocessor command 
                     (equivalent to '--cpp')
  VT_CPPFLAGS        C preprocessor flags 
                     (equivalent to '--cppflags')

build-options:
  -o, --output=PREFIX
                     Prefix of output wrapper library.
                     (default: libwrap)               

  --shared           Do only build shared wrapper library.

  --static           Do only build static wrapper library.

  --libtool=LT       Libtool command

  --cc=CC            C compiler command (default: gcc)

  --cflags=CFLAGS    C compiler flags

  --ld=LD            linker command (default: CC)

  --ldflags=LDFLAGS  linker flags, e.g. -L<lib dir>
                     (default: CFLAGS)

  --libs=LIBS        libraries to pass to the linker, 
                     e.g. -l<library>

environment variables:
  VT_CC              C compiler command 
                     (equivalent to '--cc')
  VT_CFLAGS          C compiler flags 
                     (equivalent to '--cflags')
  VT_LD              linker command 
                     (equivalent to '--ld')
  VT_LDFLAGS         linker flags 
                     (equivalent to '--ldflags')
  VT_LIBS            libraries to pass to the linker
                     (equivalent to '--libs')

examples:
  Generating wrapper library 'libm_wrap' for the Math library
  'libm.so':

    vtlibwrapgen -l libm.so -g MATH -o mwrap.c \
    /usr/include/math.h
    vtlibwrapgen --build -o libm_wrap mwrap.c
    export LD_PRELOAD=$PWD/libm_wrap.so:libvt.so


Application Execution Wrapper (vtrun) 

 vtrun - application execution wrapper for VampirTrace.

 Syntax: vtrun [options] <executable> [arguments]

   options:
     -h, --help          Show this help message.

     -V, --version       Show VampirTrace version.

     -v, --verbose       Increase output verbosity.
                         (can be used more than once)

     -q, --quiet         Enable quiet mode.
                         (only emergency output)

     -<seq|mpi|mt|hyb>   Set application's parallelization type.
                         It's only necessary if it could not 
                         be determined automatically.
                         seq = sequential
                         mpi = parallel (uses MPI)
                         mt  = parallel (uses OpenMP/POSIX threads)
                         hyb = hybrid parallel (MPI + Threads)
                         (default: automatically)

     --fortran           Set application's language to Fortran.
                         It's only necessary for MPI-applications 
                         and if it could not be determined 
                         automatically.

     --dyninst           Instrument user functions by Dyninst.

     --extra-libs=LIBS   Extra libraries to preload.

   example:
     original:
        mpirun -np 4 ./a.out
     with VampirTrace:
        mpirun -np 4 vtrun ./a.out


IOFSL server startup script (vtiofsl-start) 

 vtiofsl-start - set environment variables and start IOFSL servers.

 Syntax: vtiofsl-start [options]

   options:
     -h, --help          Show this help message.

     -V, --version       Show VampirTrace version.

     -v, --verbose       Increase output verbosity.
                         (can be used more than once)

     -q, --quiet         Enable quiet mode.
                         (only emergency output)

     -n, --num NUM       Number of IOFSL servers to start.

     -m, --mode MODE     IOFSL mode (MULTIFILE or MULTIFILE_SPLIT).
                         (default: MULTIFILE_SPLIT)

     --asyncio           Use asynchronous I/O.

   environment variables:
     VT_IOFSL_NUM_SERVERS
                         equivalent to '-n' or '--num'
     VT_IOFSL_MODE       equivalent to '-m' or '--mode'
     VT_IOFSL_ASYNC_IO=<yes|true|1>
                         equivalent to '--asyncio'

   note:
     This script needs to be sourced from a shell, since it sets
     environment variables.
     Either -n or VT_IOFSL_NUM_SERVERS must be specified.


IOFSL server shutdown script (vtiofsl-stop) 

 vtiofsl-stop - stop running IOFSL servers.

 Syntax: vtiofsl-stop [options]

   options:
     -h, --help          Show this help message.

     -V, --version       Show VampirTrace version.

     -v, --verbose       Increase output verbosity.
                         (can be used more than once)

     -q, --quiet         Enable quiet mode.
                         (only emergency output)

   note:
     This script needs to be sourced from a shell, since it sets
     environment variables.

Counter Specifications


PAPI

Available counter names can be queried with the PAPI commands papi_avail and papi_native_avail. Depending on the hardware there are limitations in the combination of different counters. To check whether your choice works properly, use the command papi_event_chooser. 

PAPI_L[1|2|3]_[D|I|T]C[M|H|A|R|W]    
              Level 1/2/3 data/instruction/total cache 
              misses/hits/accesses/reads/writes

PAPI_L[1|2|3]_[LD|ST]M    
              Level 1/2/3 load/store misses                       

PAPI_CA_SNP   Requests for a snoop                                
PAPI_CA_SHR   Requests for exclusive access to shared cache line  
PAPI_CA_CLN   Requests for exclusive access to clean cache line   
PAPI_CA_INV   Requests for cache line invalidation                
PAPI_CA_ITV   Requests for cache line intervention                

PAPI_BRU_IDL  Cycles branch units are idle                        
PAPI_FXU_IDL  Cycles integer units are idle                       
PAPI_FPU_IDL  Cycles floating point units are idle                
PAPI_LSU_IDL  Cycles load/store units are idle                    

PAPI_TLB_DM   Data translation lookaside buffer misses            
PAPI_TLB_IM   Instruction translation lookaside buffer misses     
PAPI_TLB_TL   Total translation lookaside buffer misses           

PAPI_BTAC_M   Branch target address cache misses                  
PAPI_PRF_DM   Data prefetch cache misses                          
PAPI_TLB_SD   Translation lookaside buffer shootdowns             

PAPI_CSR_FAL  Failed store conditional instructions               
PAPI_CSR_SUC  Successful store conditional instructions           
PAPI_CSR_TOT  Total store conditional instructions                

PAPI_MEM_SCY  Cycles Stalled Waiting for memory accesses          
PAPI_MEM_RCY  Cycles Stalled Waiting for memory Reads             
PAPI_MEM_WCY  Cycles Stalled Waiting for memory writes            

PAPI_STL_ICY  Cycles with no instruction issue                    
PAPI_FUL_ICY  Cycles with maximum instruction issue               
PAPI_STL_CCY  Cycles with no instructions completed               
PAPI_FUL_CCY  Cycles with maximum instructions completed          

PAPI_BR_UCN   Unconditional branch instructions                   
PAPI_BR_CN    Conditional branch instructions                     
PAPI_BR_TKN   Conditional branch instructions taken               
PAPI_BR_NTK   Conditional branch instructions not taken           
PAPI_BR_MSP   Conditional branch instructions mispredicted        
PAPI_BR_PRC   Conditional branch instructions correctly
              predicted 

PAPI_FMA_INS  FMA instructions completed                          
PAPI_TOT_IIS  Instructions issued                                 
PAPI_TOT_INS  Instructions completed                              
PAPI_INT_INS  Integer instructions                                
PAPI_FP_INS   Floating point instructions                         
PAPI_LD_INS   Load instructions                                   
PAPI_SR_INS   Store instructions                                  
PAPI_BR_INS   Branch instructions                                 
PAPI_VEC_INS  Vector/SIMD instructions                            
PAPI_LST_INS  Load/store instructions completed                   
PAPI_SYC_INS  Synchronization instructions completed              
PAPI_FML_INS  Floating point multiply instructions                
PAPI_FAD_INS  Floating point add instructions                     
PAPI_FDV_INS  Floating point divide instructions                  
PAPI_FSQ_INS  Floating point square root instructions             
PAPI_FNV_INS  Floating point inverse instructions                 

PAPI_RES_STL  Cycles stalled on any resource    
PAPI_FP_STAL  Cycles the FP unit(s) are stalled 

PAPI_FP_OPS   Floating point operations         
PAPI_TOT_CYC  Total cycles                      
PAPI_HW_INT   Hardware interrupts


CPC

Available counter names can be queried with the VampirTrace tool vtcpcavail. In addition to the counter names, it shows how many performance counters can be queried at a time. See below for a sample output. 

% ./vtcpcavail
CPU performance counter interface: UltraSPARC T2
Number of concurrently readable performance counters
on the CPU: 2

Available events:
AES_busy_cycle
AES_op
Atomics
Br_completed
Br_taken
CPU_ifetch_to_PCX
CPU_ld_to_PCX
CPU_st_to_PCX
CRC_MPA_cksum
CRC_TCPIP_cksum
DC_miss
DES_3DES_busy_cycle
DES_3DES_op
DTLB_HWTW_miss_L2
DTLB_HWTW_ref_L2
DTLB_miss
IC_miss
ITLB_HWTW_miss_L2
ITLB_HWTW_ref_L2
ITLB_miss
Idle_strands
Instr_FGU_arithmetic
Instr_cnt
Instr_ld
Instr_other
Instr_st
Instr_sw
L2_dmiss_ld
L2_imiss
MA_busy_cycle
MA_op
MD5_SHA-1_SHA-256_busy_cycle
MD5_SHA-1_SHA-256_op
MMU_ld_to_PCX
RC4_busy_cycle
RC4_op
Stream_ld_to_PCX
Stream_st_to_PCX
TLB_miss

See the "UltraSPARC T2 User's Manual" for descriptions of these
events. Documentation for Sun processors can be found at:
http://www.sun.com/processors/manuals


NEC SX Hardware Performance Counter

This is a list of all supported hardware performance counters for NEC SX machines. 

SX_CTR_STM    System timer reg
SX_CTR_USRCC  User clock counter
SX_CTR_EX     Execution counter
SX_CTR_VX     Vector execution counter
SX_CTR_VE     Vector element counter
SX_CTR_VECC   Vector execution clock counter
SX_CTR_VAREC  Vector arithmetic execution clock counter
SX_CTR_VLDEC  Vector load execution clock counter
SX_CTR_FPEC   Floating point data execution counter
SX_CTR_BCCC   Bank conflict clock counter
SX_CTR_ICMCC  Instruction cache miss clock counter
SX_CTR_OCMCC  Operand cache miss clock counter
SX_CTR_IPHCC  Instruction pipeline hold clock counter
SX_CTR_MNCCC  Memory network conflict clock counter
SX_CTR_SRACC  Shared resource access clock counter
SX_CTR_BREC   Branch execution counter
SX_CTR_BPFC   Branch prediction failure counter


Resource Usage

The list of resource usage counters can also be found in the manual page of getrusage. Note that, depending on the operating system, not all fields may be maintained. The fields supported by the Linux 2.6 kernel are shown in the table.

Name Unit Linux Description
ru_utime ms x Total amount of user time used.
ru_stime ms x Total amount of system time used.
ru_maxrss kB   Maximum resident set size.
ru_ixrss kB × s   Integral shared memory size (text segment) over the runtime.
ru_idrss kB × s   Integral data segment memory used over the runtime.
ru_isrss kB × s   Integral stack memory used over the runtime.
ru_minflt # x Number of soft page faults (i.e. those serviced by reclaiming a page from the list of pages awaiting reallocation).
ru_majflt # x Number of hard page faults (i.e. those that required I/O).
ru_nswap #   Number of times a process was swapped out of physical memory.
ru_inblock #   Number of input operations via the file system. Note: This and ru_oublock do not include operations with the cache.
ru_oublock #   Number of output operations via the file system.
ru_msgsnd #   Number of IPC messages sent.
ru_msgrcv #   Number of IPC messages received.
ru_nsignals #   Number of signals delivered.
ru_nvcsw # x Number of voluntary context switches, i.e. because the process gave up the processor before it had to (usually to wait for some resource to be available).
ru_nivcsw # x Number of involuntary context switches, i.e. a higher priority process became runnable or the current process used up its time slice.

Using VampirTrace with IOFSL

Introduction

VampirTrace and OTF can make use of the I/O Forwarding Scalability Layer (IOFSL) which allows users to write the data of many streams of a parallel trace into one or few physical files (so called multifiles) during program run. Compared with the default of writing at least two files per stream, process or even thread, this can provide a substantial performance benefit and is especially important for stability when recording highly parallel traces.

Overview

This section gives an overview over the architecture and principles from a technical point of view.

File handling in OTF

The Open Trace Format (OTF) is utilized by VampirTrace to store its trace information obtained during a run of the instrumented application. The OTF library provides an interface for reading and writing trace files. A trace consists of one or more so called streams, each containing the data of one process or thread. The data is stored in records encoded using a plain ASCII format and can optionally be transparently compressed. Although it basically offers a way to store several streams in one physical file, it does not offer mechanisms to assure data consistency for concurrent writes into one file.

To allow for arbitrary thread creation during a trace run and to avoid expensive locking, VampirTrace writes the obtained data of each process or thread into separate OTF files causing the creation of at least two files per process/thread (definitions and events). With the ever increasing number of parallel processes and the limitations of today's parallel filesystem's meta-data processing, this can become a severe problem for system performance and stability. Consequently, the goal was to significantly reduce the number of physical files used by VampirTrace and OTF during a trace run from at least two files per process/thread to a number that is acceptable for today's filesystems.

I/O Forwarding Scalability Layer

The goal of the I/O Forwarding Scalability Layer IOFSL is to provide a forwarding layer on the basis of a client-server architecture. It allows clients to send I/O requests to a server which is able to execute the original I/O calls and even aggregate these requests to improve performance. Besides the aggregation of normal write requests, the server also offers non-blocking write requests and a so-called atomic append mode which allows many clients to write potentially large blocks of data concurrently into one single physical file (multifile) without the need for client-side locking. In this case, the data is appended to the end of the file and the corresponding offset can be obtained later. Additionally, this atomic append feature can be used with more than one server allowing the write requests of many clients into one file being distributed across a smaller number of servers.

IOFSL is being developed at Argonne National Laboratory and is available at www.iofsl.org. By relying on open software, it is portable to a wide range of machines and has been tested on a generic Linux cluster as well as on the leadership-class computing system Jaguar.

Architecture

Integrating the three previously described parts leads to an architecture with VampirTrace and OTF built on top of IOFSL. The instrumented application generates events that are handled and buffered by the VampirTrace runtime library. When the thread local buffer is full, the events are passed to the OTF library where they are compressed. If the IOFSL mode is enabled, the resulting write buffers are passed to the IOFSL client library (zoidfs) which sends the data to the IO forwarding servers where it is aggregated (atomic append), buffered and finally sent out to the file system.

Since IOFSL servers can handle multiple clients, an N:M mapping of clients to servers is possible. The exact ratio depends on the amount of data the clients send and the bandwidth available for the server nodes. In our test cases, a ratio of up to 300 clients per server was used.

When using the IOFSL integration, all write requests in OTF are issued using the zoidfs API[*]. Those writes are handled by the IOFSL forwarding servers and aggregated into a single file using the atomic append feature. The offset in the multifile is returned to OTF and stored in a second file, the so called index file, in order to maintain the mapping between written blocks and streams. For any block of a stream written into the multifile, the index file contains the ID of the stream, the start of the block, and its length. This allows for an efficient reading of blocks since only the index file has to be scanned for entries for a given stream ID. Additionally, a large number of logical files (streams) can be stored using only two physical files.

Installation

In order to use this setup, IOFSL and VampirTrace have to be compiled in order. In the following sections, the directory <install_dir> should be replaced with a - possibly user-local - directory used for installation, e.g. $HOME/local[*]. The installation procedure for IOFSL is described at https://trac.mcs.anl.gov/projects/iofsl/wiki/Building. Currently the iofsl_vampir git branch is required.

Support Libraries

IOFSL requires several libraries in order to work correctly:

Note that building boost, OpenPA or BMI/PVFS is not required in case it is already present on the machine. Building GNU autoconf is not covered by this document. For the use with VampirTrace, ROMIO and therefore rebuilding MPICH is not required.

Building Boost

Boost Version 1.46.1 is recommended, other Versions might be incompatible. To build the required boost libraries, issue the following commands in the source directory: 
$> ./bootstrap.sh \
   --with-libraries=system,date_time,\
program_options,regex,thread,test \
   --prefix=<install_dir>

$> ./bjam  --prefix=<install_dir> \
   --libdir=<install_dir>/lib \
   --includedir=<install_dir>/include \
   install

Building OpenPA

To build the required OpenPA library, issue the following commands in the source directory: 

$> ./configure --prefix=<install_dir>
$> make all install

Building BMI/PVFS

To build the required BMI/PVFS library, issue the following commands in the source directory: 

$> ./configure --enable-bmi-only --prefix=<install_dir> \
   --with-openib=<openib_install_dir>
$> make all install
Note that the option -with-openib can be omitted if support for direct access to InfiniBand is not required.

Building IOFSL

Create a local copy of the git reposotiry branch: 

$> mkdir iofsl
$> cd iofsl
$> git init
$> git remote add -t iofsl_vampir \
   -f origin git://git.mcs.anl.gov/iofsl.git
$> git checkout iofsl_vampir
$> ./prepare

The following commands can be used to build the IOFSL client and server: 

$> ./configure --with-bmi=<install_dir> \
   --with-boost=<install_dir> --with-openpa=<install_dir> \
   --prefix=<install_dir> --with-cunit=no

$> make all install

Building VampirTrace & OTF

After extracting the source code from the archive, issue the following commands: 
$> ./configure                         \
   --prefix=<install_dir>              \
   --enable-iofsl                      \
   --with-zoidfs-dir=<install_dir>     \
   --with-bmi-dir=<install_dir>        \
# On Cray XK6 with PBS as batch system add
   --enable-iofsl-scripts=crayxk6
$> make all install


Usage Examples

The use of I/O forwarding servers implicates a system specific deployment. VampirTrace mitigates this effort by providing convenient scripts for specific system setups. Currently Cray XK6 systems are supported, which are described here. Furthermore the IOFSL specific adjustable parameters of VampirTrace are described.

Using VampirTrace with IOFSL on Cray XK6 / with PBS

Building your application with VampirTrace

We assume that VampirTrace with IOFSL support has been installed as previously described. This might be deployed to the user using a module. 

   # Check module av vampirtrace to
   # see what is available at your system
$> module load vampirtrace/5.13

Build your application as usual with VampirTrace. For details please refer to the general part of this documentation. 

$> vtcc -vt:hyb application.c -o application

Running an Example

The scripts vtiofsl-start and vtiofsl-stop are provided to control the IOFSL server instances. They will be launched on dedicated compute nodes that are part of the batch Job allocation.

PBS Options

It is important to reserve a sufficient number of processor cores. The number of cores requested must be large enough to contain the number of application cores plus the number of cores required for the IOFSL server instances. Each IOFSL server will run on a dedicated node[*].Thus N_allocated ≥((N_IOFSL * 16) + N_Application) must hold.

Example using 64 server instances: 

#!/bin/sh
#PBS...
[...]
## Allocate enough cores: (64 * 16) + 16384 => 17408
#PBS -l size=17408
## Preserve environment
#PBS -V

Environment Variables

It is highly recommended to set the following environment variable.

Example: 
[...]
# The directory to which the trace is written
mkdir trace
export VT_PFORM_GDIR=$PWD/trace

Execution

Launching and stopping the servers as is done using the supplied scripts. The scripts are sourced from the job script or interactive shell to allow them setting required environment variables for VampirTrace. 
[...]
# rca module need to be loaded!
. /opt/modules/default/etc/modules.sh
module load rca

# Start server
source vtiofsl-start -n 64

# Run application as usual
aprun -n 16384 application --parameter inputfile

# Shutdown server
source vtiofsl-stop

Interactive Jobs

Interactive jobs work the same way. You can either run a script similar to the job submission script, or run the commands from your shell. However the scripts are developed and tested on bash. Other shells are not supported.

The vtiofsl-scripts assume to be run within a PBS job. If you run them multiple times within one job, the detailed log files may be overwritten.

Log files and debug information

The vtiofsl-scripts create a number of log files and configuration files in the $VT_PFORM_GDIR/.iofsl directory.


Manual Usage

The machine specific installation strives to hide most of the complexity of the I/O forwarding solution from the end-user. In the background, the forwarding server(s) are started and environment variables are set in order to point VampirTrace / OTF to them.

Configuring the Server

The server is configured using a configuration file. At server start-up, this file is provided using the -config argument. The cray XK6 configuration file is provided in the package[*]. For more information about the options available please refer to the IOFSL documentation[*]. The most important option is the serverlist entry in the bmi section which takes a list of server addresses, e.g. : 
bmi
{
  serverlist =  ( "tcp://192.168.97.236:12345", 
                  "tcp://192.168.97.237:12345", 
                  "tcp://192.168.97.238:12346" );
}
At start-up, the server looks for the environment variable ZOIDFS_SEVER_RANK to determine its address, e.g. ZOIDFS_SEVER_RANK=0 would cause the address tcp://192.168.97.236:12345 to be used. The configuration file can be shared between all server instances and lets the servers determine the coordination server, which is usually rank 0.

Launching the Servers

The I/O forwarding server (iofwd) can be deployed in multiple ways. This is highly system specific, possible ways to do so are:

Pointing VampirTrace to the servers

The list of available I/O forwarding servers is provided to VampirTrace by setting VT_IOFSL_SERVERS to a comma-separated list of addresses, e.g. 

export VT_IOFSL_SERVERS= \
     "tcp://192.168.1.1:12345,tcp://192.168.1.2:12345"
VampirTrace / OTF will choose a server upon opening the file based on the stream identifier encoded in the original filename.

File modes

In the default setting, each server will create two files for each type of file, the actual file containing the appended data and an index file. This mode is called MULTIFILE_SPLIT. It provides a good workload for parallel file systems. In the so called MULTIFILE mode, all servers share data and index files. It requires additional synchronization between the servers. Also the Lustre file system does not allow to stripe individual files over more than a maximum number of storage targets, introducing a performance-bottleneck. The MULTIFILE mode should be considered experimental. Therefore, using the default mode is recommended. The mode can be set using VT_IOFSL_MODE to either MULTIFILE_SPLIT or MULTIFILE.

Asynchronous I/O

IOFSL offers a capability, where write requests are buffered on the forwarding server. This can reduce the trace flush times, without consuming node local resources. To enable this, VT_IOFSL_ASYNC_IO is set to yes.

Unification

The unification step can also use the IOFSL mode for writing the output trace. This is controlled with the same environment variables. Therefore if VampirTrace uses IOFSL, the implicit unification at the end of the trace run will also use IOFSL for output. If VT_UNIFY=no, then one should make sure that the correct IOFSL environment is also available to the later vtunify(-mpi), unless intended otherwise.

Compatibility of the generated trace

All tools that work on the generated trace need to be built with the appropriate OTF Version to ensure compatibility with traces generated with IOFSL. This especially applies to the Vampir visualization server and GUI. If backwards compatibility is required, the trace can be transformed using otfmerge, e.g. 

$> mpirun -np 1024 \
       otfmerge-mpi -n 0 -o merged-trace input-trace.otf

FAQ


Can I use different compilers for VampirTrace and my application?

There are several limitations which make this generally a bad idea:

For example, the combination of a GCC compiled VampirTrace with an Intel compiled application will work except for OpenMP. But to avoid any trouble it is advisable to compile both VampirTrace and the application with the same compiler.


Why does my application need such a long time for starting?

If subroutines have been instrumented with automatic instrumentation by GNU, Intel, PathScale, or Open64 compilers, VampirTrace needs to look-up the function names and their source code line before program start. In certain cases, this may take very long. To accelerate this process prepare a file with symbol information using the command nm as explained in Section 2.3 and set VT_GNU_NMFILE to the pathname of this file. This method prevents VampirTrace from getting the function names from the binary.


How can I limit compiler instrumentation?

Fully-automatic instrumentation by the compilers is the most convenient method to instrument your program. However, a variety of functions will be instrumented and all calls of these functions will be traced. Runtime filters do not eliminate complete overhead of tracing automatically instrumented functions. Therefore, it is often desirable to limit compiler instrumentation to specific functions. Several compilers provide options to configure function instrumentation. Start with VampirTrace in Profiling Mode by setting VT_MODE to STAT. The profiling information can be used to determine functions which may be excluded from automatic instrumentation.

The IBM C compiler ≥11 and Fortran compiler ≥13 provide -qfunctrace option to enable tracing for all functions. To disable tracing for all functions you can use -qnofunctrace. Regardless of -qnofunctrace both -qfunctrace+ and -qfunctrace- can be used to enable resp. disable tracing for a colon-separated list of function names, classes, or namespaces. For example,



-qfunctrace -qfunctrace-myFunc1:myFunc2


enables tracing for all functions except for myFunc1 and myFunc2.

Also GNU compiler ≥4.3 provides options to limit compiler instrumentation. -finstrument-functions-exclude-file-list sets a list of files. All functions defined in a file of this list will be excluded from instrumentation. The option -finstrument-functions-exclude-function-list sets a list of function names that are excluded from instrumentation. Arguments of both compiler options must be separated by comma. Matching of arguments with function or file names is done on substrings. For example,



-finstrument-functions-exclude-file-list=include


will exclude any function defined in files whose pathnames contain "include". Maybe such a rule is too restrictive, because the "include" directory of your own program code is affected too. The pattern needs to be specified more precisely, for instance:



-finstrument-functions-exclude-file-list=/usr/include


This rule can be used to exclude Standard Template Library (STL) calls in C++ from tracing.


Why do I see multiple I/O operations for a single (un)formatted file read/write from my Fortran application?

VampirTrace does not implement any tracing at the Fortran language level. Therefore it is unaware of any I/O function calls done by Fortran applications.

However, if you enable I/O tracing using VT_IOTRACE, VampirTrace records all calls to LIBC's I/O functions. As Fortran uses the LIBC interface for executing its I/O operations, these function calls will be part of the trace. Depending on your Fortran compiler, a single Fortran file read/write operation may be split into several LIBC read calls which you will then see in your trace.

Beware that this may lead you to the (wrong) conclusion that your application spends time between the LIBC I/O calls inside the user function that contains the Fortran I/O call, especially when doing formatted I/O . It is rather the Fortran I/O subsystem which does all the formatting of the data that is eating your cpu cycles. But as this layer is unknown to VampirTrace, it cannot be shown and the time is accounted to the next higher function in the call stack - the user function.

The application has run to completion, but there is no *.otf file. What can I do?

The absence of an *.otf file usually means that the trace was not unified. This is the case on certain platforms, e.g. when using DYNINST or when the local traces are not available when the application ends and VampirTrace performs trace unification.

In those cases, a *.uctl file can be found in the directory of the trace file and the user needs to perform trace unification manually.


What limitations are associated with "on/off" and buffer rewind?

Starting and stopping tracing by using the VT_ON/VT_OFF calls as well as the buffer rewind method are considered advanced usage of VampirTrace and should be performed with care. When restarting the recording of events, the call stack of the application has to have the same depth as when the recording was stopped. The same applies for the rewind call, which has to be at the same stack level as the rewind mark. If this is not the case, an error message will be printed during runtime and VampirTrace will abort execution. A safe method is to call VT_OFF and VT_ON in the same function.

It is allowed to use "on/off" in a section between a rewind mark and a buffer rewind call. But it is not allowed to call VT_SET_REWIND_MARK or VT_REWIND during a section deactivated by the "on/off" functionality.

Buffer flushes interfere with the rewind method: If the trace buffer is flushed after the call to VT_SET_REWIND_MARK, the mark is removed and a subsequent call to VT_REWIND will not work and issue a warning message.

In addition, stopping or rewinding tracing while waiting for MPI messages can cause those MPI messages not to be recorded in the trace. This can cause problems when analyzing the OTF trace afterwards, e.g.,  with Vampir.


VampirTrace warns that it ``cannot lock file a.lock'', what's wrong?

For unique naming of multiple trace files in the same directory, a file *.lock is created and locked for exclusive access if VT_FILE_UNIQUE is set to yes (⇒ Section 3.1). Some file systems do not implement file locking. In this case, VampirTrace still tries to name the trace files uniquely, but this may fail in certain cases. Alternatively, you can manually control the unique file naming by setting VT_FILE_UNIQUE to a different numerical ID for each program run.


Can I relocate my VampirTrace installation without rebuilding from source?

VampirTrace hard-codes some directory paths in its executables and libraries based on installation paths specified by the configure script. However, it's possible to move an existing VampirTrace installation to another location and use it without rebuild from source. Therefore it's necessary to set the environment variable VT_PREFIX to the new installation prefix before using VampirTrace's Compiler Wrappers (⇒ Section 2.1) or launching an instrumented application. For example: 

./configure --prefix=/opt/vampirtrace
make install
mv /opt/vampirtrace $HOME/vampirtrace
export VT_PREFIX=$HOME/vampirtrace


What are the byte counts in collective communication records?

The byte counts in collective communication records changed with version 5.10.

From 5.10 on, the byte counts of collective communication records show the bytes per rank given to the MPI call or returned by the MPI call. This is the MPI API perspective. It is next to impossible to find out how many bytes are actually sent or received during a collective operation by any other MPI implementation.

In the past (until VampirTrace version 5.9), the byte count in collective operation records was defined differently. It used a simple and naive hypothetical implementation of collectives based on point-to-point messages and derived the byte counts from that. This might have been more confusing than helpful and was therefore changed.

Thanks to Eugene Loh for pointing this out!


I get ``error: unknown asm constraint letter''

It is a known issue with the tau_instrumentor that it doesn't support inline assembler code. At the moment there is no other solution than using another kind of instrumentation like compiler instrumenation (⇒ Section 2.3) or manual instrumenation (⇒ Section 2.4).


I have a question that is not answered in this document!

You may contact us at mailto:vampirsupport@zih.tu-dresden.devampirsupport@zih.tu-dresden.de for support on installing and using VampirTrace.


I need support for additional features so I can trace application xyz.

Suggestions are always welcome (contact: vampirsupport@zih.tu-dresden.de) but there is a chance that we can not implement all your wishes as our resources are limited.

Anyways, the source code of VampirTrace is open to everybody so you may implement support for new stuff yourself. If you provide us with your additions afterwards we will consider merging them into the official VampirTrace package.

Footnotes

... (OTF)[*]
http://www.tu-dresden.de/zih/otf
... tool [*]
http://www.vampir.eu
... Open MPI [*]
http://www.open-mpi.org/faq/?category=vampirtrace
... documentation [*]
http://www.cs.uoregon.edu/Research/tau/docs/newguide/bk05ch02.html#d0e3770
... Dyninst [*]
http://www.dyninst.org
... library [*]
http://sourceforge.net/projects/ctool
... CLAPACK[*]
www.netlib.org/clapack
... Dyninst [*]
http://www.dyninst.org
... PDToolkit [*]
http://www.cs.uoregon.edu/research/pdt/home.php
... TAU [*]
http://tau.uoregon.edu
... CTool [*]
http://sourceforge.net/projects/ctool
... API[*]
The OTF master control file is written using POSIX I/O in any case.
...$HOME/local[*]
The software packages can be installed in different directories.
... node[*]
The server makes use of all the nodes resources by multithreading and allocating large I/O buffers
... package[*]
tools/vtiofsl/platform/crayxk6-iofwd.cf
... documentation[*]
https://trac.mcs.anl.gov/projects/iofsl/wiki/ConfigurationFile