profiling-intel

This repository houses the course materials for the ARC training course: "A brief introduction to profiling with the Intel toolchain".

Slides can be found in slides/. Materials for excercises can be found in auto_tracing/ and user_instrumentation/. These excercises are designed to be run on the Hamilton 8 supercomputer at Durham, but will run on other machines that offer Intel MPI with only a few tweaks to the job submission scripts.

Excercise 1

Prerequisites:

• ssh into Hamilton 8.
• clone this repository
• Change into the auto_tracing/ directory

Have a look at auto_tracing/itac_test.cpp. The code is strange (it doesn't really do any proper computations), but it allows us to experiment with ITAC. Notice:

• the code is intended to be run with 2 ranks and is not multithreaded
• Rank 0 makes 2 nonblocking sends to rank 1 before doing some physics
• Rank 1 intersperses physics with receiving the messages
• Processes are blocked at the end of the programme until all processes reach the barrier

Tasks:

1. Submit auto_tracing/submit.sh to Hamilton (should run very quickly)
2. Open the resulting trace file with traceanalyzer
3. Open the event timeline
4. Ungroup the MPI functions and locate the call to MPI_Barrier
5. Filter the communications so that you only see messages sent with MPI tag 0

Exercise 2 (extension)

Change into the user_instrumentation/ directory. Submit the code to Hamilton using submit.sh. Open the event timeline and ungroup the functions to view the user defined instrumentation.