Dear Robert,
Adding to the comments and hints by Lorenzo and Paolo, I would just like
to add that i) I usually use OpenBLAS (https://www.OpenBLAS.net/) instead
of the NetLib versions of BLAS+LAPACK (compiled from source, it easily
adapts to the underlying hardware, with few options and it seems to me to
be close to the performance of Intel MKL et co; despite its name it also
contains LAPACK, not only BLAS) and ii) why OpenMPI 1.x.y, as the latest
versions are at 4.x.y?
I would also ensure that the environmental variable 'OMP_NUM_THREADS' is
indeed set correctly, even if at '1', to avoid a burst of OpenMP threads.
The correct number of course depends on the number of cores available,
hyper-threading or not (or how is it called on the AMD EPYC), the number
of MPI tasks and the amount of memory available, should the latter become
a bottleneck.
Greetings from Paris,
apsi
-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-
Ari Paavo Seitsonen / [email protected] / http://www.iki.fi/~apsi/
On Sat, 23 Jan 2021, Robert Stanton wrote:
Dear all,
I am able to successfully install parallel QE and get everything to run
smoothly, however after testing against an install on an
externally maintained cluster I'm finding my speed to be slower by a factor of
5-10x when keeping the system, core number, QE version,
etc. the same.
I'm not sure the degree to which this could be a hardware discrepancy (Xeon
Silvers vs new EPYCs), v.s. how much could be a
sub-optimal compilation of QE. The slower QE install is being done with
blas/lapack 3.8.0 and openmpi 1.10.7. I'm just curious if
there is anything very clearly not ideal about this environment, or if it is
likely something I'll just need to troubleshoot with
smaller settings? Thanks so much in advance for any insights you can provide!
Regards,
Robert Stanton
Graduate Student
Clarkson University
_______________________________________________
Quantum ESPRESSO is supported by MaX (www.max-centre.eu)
users mailing list [email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
