libiomp5md.dll performance issue

[cabrita__christophe]

Hello,

We are experiencing a severe performance degradation when upgrading from libiomp5md.dll V5.0.2017.706 (delivered with IntelFortran V2017.5) to libiomp5md.dll V5.0.2018.913 (Fortran V2019.1).

Scenario:

- 2 instances of a multithread 64-bit executable running simultaneously on Windows7 Pro (V6.1 no 7601 : SP1)

- output from OpenMp KMP_AFFINITY=none,verbose:

OMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.
OMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info
OMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 0-5
OMP: Info #156: KMP_AFFINITY: 6 available OS procs
OMP: Info #157: KMP_AFFINITY: Uniform topology
OMP: Info #179: KMP_AFFINITY: 1 packages x 6 cores/pkg x 1 threads/core (6 total cores)

- Launching 2 instances of the executable with 6 threads per process provokes very slow execution for both processes (approx. 50 times slower than normal). Killing either one of the processes immediately "accelerates" the remaining process.

- Modifying the number of threads per process has the following effect:

- 5 threads: both processes execute rapidly

- 7 threads: both processes execute rapidly, though not quite as rapidly as with 5 threads

- Replacing libiomp5md.dll V5.0.2018.913 with libiomp5md.dll V5.0.2017.706 (no rebuilding of the executables) results in rapid execution is all cases.

I have not yet produced a simple minimal executable that reproduces the problem though I will obviously try to if necessary. Before I set about doing so can the Intel support team say if they are any known issues with libiomp5md.dll V5.0.2018.913? Can I supply any more info to help with the diagnosis?

Regards,

John Sturton

[Barbara_P_Intel]

Try the current release, 19.0 update 3.  There is fix for the performance degradation of running multiple OpenMP jobs simultaneously.

[cabrita__christophe]

Thanks, I will try the 19.0U3 version. Is this fix mentioned in any of the release notes?

[Barbara_P_Intel]

It's not in the release notes.

[cabrita__christophe]

The 19.0U3 version of libiomp5md.dll (V5.0.2019.109) seems to fix the problem - we now have execution times comparable to those observed when using libiomp5md.dll V5.0.2017.706. Thank you for your help.

For info: We had previously tried with the version of libiomp5md.dll delivered with 2019Update2 but this did not change anything , and since the release notes for 2019Update3 did not mention any major modifications to OpenMP we did not try the 2019Update3 version of libiomp5md.dll. Surely a major fix like this is worth shouting about?

Regards

John Sturton

[andrew_4619]

cabrita, christophe wrote:
....and since the release notes for 2019Update3 did not mention any major modifications to OpenMP we did not try the 2019Update3 version of libiomp5md.dll. Surely a major fix like this is worth shouting about?

That is an issue that many users have complained about. You find a problem but get no visibility of when it is fixed. It used to be that if there was a thread discussing the problem the intel staff would at least add a note to it when that got fixed but even that does not happen any more and in fact intel only act on stuff that has a support ticket filed. Support issues have ZERO visibility for other users other than the person that files the ticket. An area of negative progress I am sorry to say.