- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -

Vampirtrace

From HLRS Platforms
The VampirTrace library allows MPI communication events of a parallel program to be recorded as a trace file. Additionally, certain program-specific events can also be included. To record MPI communication events, simply relink the program with the VampirTrace library. A new compilation of the program source code is only necessary if program-specific events should be added.
Developer: ZIH TU Dresden and JSC/FZ Juelich
Platforms: NEC Nehalem Cluster
Category: Performance Analyzer
License: BSD License
Website: VampirTrace homepage


Introduction

Tracing an application with Vampir is done in three steps:

  1. Compiling with the vampirtrace wrappercompiler
  2. Running the application
  3. Analysing the trace files using vampir or vampir-server

For detailed usage of Vampirtrace see the user manual which can either be found on your local machine or in the internet.

Usage

Compiling with the vampirtrace wrappercompiler

First load the needed software module:

module load performance/vampirtrace

Now you can compile your application using the wrapper compilers vtcc, vtc++ or vtf90 as described in the user VampirTrace manual.

Note: When using vampirtrace in combination with MPI you will need at least to specify the name of your mpi compiler to VampirTrace
vtf90 -vt:f90 mpif90
vtcc  -vt:cc  mpicc
vtcxx -vt:cxx mpicxx


You can specify the amount of instrumentation to be include in the trace using -vt options, e.g. -vt:mpi for MPI instrumentation, -vt:compinst for compiler-based function instrumentation:

vtcc -vt:mpi -vt:inst compinst -o myapp myapp.c


Hermit special:

Warning: On Hermit it is necessary to add the -vt:mpi option. Otherwise the wrong vampirtrace libraries will be used.
vtcc -vt:mpi [other options] my_app

BW-Grid special:

On the BW-Grid you first have to login to a node for compilation as some libraries are not accessible from the frontend:

qsub -I -l walltime=1:00:00


Generating the trace files

Run your application as usual. This will generate the needed trace files.

Which data are collected during the run can be controlled with the VT_* options. Options which may be helpful during analysis are VT_BUFFER_SIZE, VT_MAX_FLUSHES, VT_PFORM_LDIR or VT_METRICS. See the Vampirtrace documentation for more details.

Viewing traces

To analyze the generated traces you can use Vampir. For large traces (>500MB), e.g. generated with more than 16 PEs in the application run, it is recommended to use VampirServer.


Trouble shooting

In this section common errors and solutions to overcome the problems are listed.

Incomplete trace files

[0]VampirTrace: Maximum number of buffer flushes reached (VT_MAX_FLUSHES=1)
[0]VampirTrace: Tracing switched off permanently

Trace data did not fit into a single VampirTrace buffer. Unlimit the number of buffer reuses with

export VT_MAX_FLUSHES=0


Not enough space on /tmp

[0]VampirTrace: FATAL: OTF_WStream_writeLeave failed:
ERROR in function OTF_WBuffer_writeUint64, ...
OTF_WBuffer_guarantee() failed.

Check if /tmp has filled up to 100% by executing

df -h | grep /tmp

A solution is to use a workspace in combination with the VT_PROFM_LDIR variable:

export VT_PROFM_LDIR=$(ws_allocate vt_trace_tmp 5)


Library issues on NEC

Try to export LD_LIBRARY_PATH

export LD_LIBRARY_PATH=/opt/NECcomp/compiler100/ia64/lib


See also

External Links