01:54 <ste||ar-github> [hpx] khuck pushed 1 new commit to khuck-patch-2: https://github.com/STEllAR-GROUP/hpx/commit/6c362d7c1417fab101ec76a476fedcfe93de17f2

01:54 <ste||ar-github> hpx/khuck-patch-2 6c362d7 Kevin Huck: Merge branch 'master' into khuck-patch-2

05:09 <ste||ar-github> [hpx] sithhell reopened pull request #3439: Use correct MPI CXX libraries for MPI parcelport (master...fix-ubuntu-bionic-mpi) https://github.com/STEllAR-GROUP/hpx/pull/3439

10:16 <_diers_> Hello, I have a problem with NUMA on a dual-socket system (2x12 cores + HT). I'm testing a 2D stencil based on the HPX tutorial. Once the two NUMA domains are used, the performance breaks down completely. Maybe someone can give me support / help to investigate the problem?

10:18 <heller> hi _diers_

10:18 <heller> first, it would help if you can share the code

10:18 <heller> second, how you run it

10:21 <_diers_> hi @heller

10:23 <_diers_> I think i can reproduce it with the unchanged tuturial code stencil_parallel_1.cpp . I will compile it.

10:28 <heller> stencil_parallel_1.cpp is not NUMA aware ;)

10:29 <heller> crap, it should be ... my mistake

10:36 <_diers_> Yes, I can reproduce it. Where should I paste the results and the call?

10:36 <heller> gist.github.com maybe?

10:49 <_diers_> Here is the result: https://gist.github.com/0dnb0/b2f6d2f36abd364ce5046dff2c8cabd8

10:52 <heller> interesting...

10:52 <heller> may I ask what "lstopo" is giving you for your allocation?

10:53 <heller> the other thing is, that the problem size seems to be very small

10:53 <_diers_> likwid is also available

10:59 <_diers_> i added a comment

11:00 <heller> lstopo would have been more interesting ;)

11:00 <heller> let me check with my local system

11:02 <_diers_> yes, but I also get almost the same behavior with a 25 point stencil.

11:04 <_diers_> It is also possible that something is wrongly configured on the system.

11:05 <heller> no, I can reproduce the results

11:05 <zao> Ooh.

11:05 <heller> it gets better once you increase the number of elements in the y direction

11:06 <heller> so, the tutorial code only parallelizes in y direction, IIRC

11:07 <_diers_> is the output of lstopo-no-graphics enough?

11:07 <heller> so it looks like there just isn't enough parallelism for more than 2 NUMA domains

11:07 <heller> yes

11:08 <_diers_> added the output of lstopo

11:09 <_diers_> Should I switch Nx,Ny or increase it?

11:10 <_diers_> ...for more than 1 NUMA domain?

11:11 <heller> if you increase Ny, you should see the expected performance behavior

11:15 <_diers_> I added a comment with an increased Ny

11:28 <heller> _diers_: that's better ;D

12:30 <_diers_> ok, but that's not a solution for me ;-)

12:43 <jbjnr> _diers_: I have been working on improved numa support and I will take a look at the stencil example. Could you please file an issue on github that describes the problem and your findings.

12:49 <heller> _diers_: ok, I guess you need the fixed size there? May I ask you to run stencil_parallel_4.cpp and see if that exhibits the same behavior?

12:53 <_diers_> @jbjnr ok i will create an issue tomorrow

12:54 <_diers_> @heller ok, i test it

13:16 <_diers_> @heller Behave the same with stencil_parallel_4.cpp. Added the results to the existing gist.

13:19 <heller> hmn, ok

18:20 <hkaiser> helyt?

18:20 <hkaiser> heller: yt?