The recent proposed patch for balanced affinity on AArch64 or other platforms - Does anyone have a benchmark which demonstrates the problem? I have access to most of the available AArch64 hw and would like to test/validate it.
This CR only has a patch and no test
https://reviews.llvm.org/D22365
Hi Chris,
Actually any test with openmp parallel region can be used for testing of this feature. I used simple “hello” test with addition of empty parallel.
What matters here:
setting of KMP_AFFINITY=balanced,verbose (verbose - to see how threads are bound, otherwise more complicated test is needed, probably different for each particular system);
running on various HW topologies with various thread numbers, and see how threads are bound. I used taskset on regular Linux Intel64 multi socket multi code HT-enabled machine to emulate various topologies.
If you meant automatic test for regular runs on different architectures, - this is problematic, because it is hard to check that threads are bound in some particular way without knowing details of a particular machine and OS. I failed to figure out how to write such a generic test, though maybe it is possible somehow.
Regards,
Andrey