pertalde 0.1.5

Some command-line utilities on paraver traces.

pertalde

pertalde is a set of tools to manipulate Paraver traces. Paraver is a parallel performance analysis tool by the Performance Tools team at BSC, and is a parallel trace visualization system allowing for large scale trace execution analysis. Paraver can be obtained at https://tools.bsc.es/downloads.

How to install

pertalde is distributed as a PyPI package and thus can be installed with pip.

python -m pip install pertalde

Suggestions for HPC clusters

Create a python environment

python -m venv ~/.pertalde_env
source ~/.pertalde_env/bin/activate

python -m pip install pertalde

When you don't have internet access

If you want to install pertalde in a remote machine where you don't have internet access, you can first download the package.

python -m pip download pertalde

This will download the package (a .whl file) which you can the install using

python -m pip install <package_name>.whl

Utilities

pertalde unir (i.e. parajuntar)

When you a trace with sparse semantic values, and/or with inconsistent value per semantic across traces you can use pertalde unir. This will use the values from the pcf files to unify the values based on their semantic, and translate the prv files accordingly. You can also group different event types together to unify their semantics.

How to use it.

If you had a trace with different event that share semantics but differ in values. For a more specific example, lets say we have a trace.prv with events 1000, 1001, 2000, 2001. Where events 1000 and 1001 share meaning between them, and 2000 and 2001 between them.

ptd unir --types=1000,1001 --types=2000,2001 trace.prv

This would result in a trace.unir.prv file where now all values that have the same name for 1000 and 1001 have also the same value. And all the same for values 2000 and 2001.

If you have multiple traces with the same event but this event has inconsistent values for the same semantic across traces, you can also use pertalde unir. By running the same command on different traces, the resulting event values will be common across traces also.

ptd unir --types=1000,1001 trace1.prv trace2.prv

A practical example that combines these two modes is when you have multiple traces with OpenMP, and you want to compare the values among these traces:

ptd unir                        \
    --sufix unified_openmp      \
    --sort-values-by-name       \ # Keep similar names closer in value
    -t 60000018                 \ # Executed parallel function
    -t 60000118                 \ # Executed parallel function line and file
    -t 60000023                 \ # Executed OpenMP task function
    -t 60000123                 \ # Executed OpenMP task function line and file
    -t 60000024                 \ # Instantiated OpenMP task function
    -t 60000124                 \ # Instantiated OpenMP task function line and file
    -c                          \ # Compress the resulting trace
    trace1.prv trace2.prv       ; # Or more traces if necessary

In the case you get a new trace and you want it to have common values with a trace you have already processed earlier you can make use of the option --base-pcf.

# This produces trace1.unir.prv trace1.unir.pcf and trace1.unir.row
ptd unir --types=1000,1001 trace1.prv

# Now you get a new trace: trace2.prv
ptd unir --types=1000,1001 --base-pcf=trace1.unir.pcf trace2.prv

This will make event values from types 1000 and 1001, form trace2.prv, consistent with those in trace1.prv.

This tool has preset gorups for specific tools. As this tool was initially implemented to solve a common problem with neSmiK and nsys2prv we have presets for this tools. For example, for nesmik you could use:

ptd unir --profile=nesmik trace.prv

pertalde separar

When you have a big trace that you want to analyse its common practice to filter it. This tool has a preset of filters that you can use without having to rely on a configuration generated from Paraver. Furthermore, it allows to run multiple filters on a single execution. Which results in generating the filters faster. Right now we have event filters for neSmiK, MPI, OpenMP, Counters, and Flushing. And we have a useful filter with a minimum duration threshold.

How to use it.

ptd separar --compress --useful 100 --profile=mpi,counters trace.prv

This will generate a directory names test.partials, which will contain 3 different traces with names trace.{useful,mpi,counters}.{prv.gz,pcf,row}. Notice that this tool can also leverage processing the trace to compress it with the flag --compress which is available for all subcommands that produce output traces.

pertalde comptar

This tool will count the occurences and estimate the uncompressed disk size of different records.

ptd comptar --all trace.prv

pertalde sincronitzar

If you get a trace that is not synchronized, you can use this tool to synchronize it based on an event. This is helpful when the merging process is very long, or you no longer have the unmerged files. This tool can only syncronize the MPI tasks.

# By default it will use the MPI init event.
ptd syncronitzar trace.prv

pertalde comprovar

This tool basically will test a trace to check if its well formatted. If you are having problems with a trace, or are generating trace files, this tool can tell you where the trace has format issues.

ptd comprovar trace.prv