Profiling Code on the Cluster

The cluster has a special dedicated partition for benchmarking, called bench. It currently contains four machines, 533,533,534 and 535, which are configured for profiling.

Before using it, you need to ask to be added to the special benchmarking account. Once you have done this, you can get access to the partition via:

• ssh -X maestro

Note -X , to get access to graphical user interface. Once you are on maestro-submit

• srun -p bench --qos bench -A bench --pty --x11 bash

We have a licence to a highly advanced profiler, Intel Vtune

• module load icc/xe_2019
• amplxe-gui

If for some reason you do not want to use the GUI version, there are other tools available:

Code Block (text)

amplxe-androidreg.sh amplxe-perf amplxe-python-execamplxe-cl amplxe-perf-sync amplxe-runssamplxe-feedback amplxe-pgo-report.sh amplxe-runss.pyamplxe-gui amplxe-python amplxe-self-checker.sh

here is the overview of Intel Vtune:

https://www.youtube.com/watch?v=Vl19kyt-cVI

or read the full documentation here:

https://software.intel.com/en-us/vtune-amplifier-help

Ask us if you need any help (ask-hpc@pasteur.fr)

for admins:

echo 1 > /proc/sys/kernel/perf_event_paranoid