[CP2K-user] [CP2K:13553] ssmp and CPU loads

[hut... at chem.uzh.ch]

Hi

you get the most flexibility and best performance when using the
PSMP version of the code. The optimal combination of MPI ranks
and OpenMP threads will depend on your hardware and input.
My first guess is usually N MPI ranks and 1 OpenMP thread.

Your memory problem might be related to overhead (within CP2K)
from the SSMP code.

regards

Juerg Hutter
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut... at chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp... at googlegroups.com wrote: -----
To: "cp2k" <cp... at googlegroups.com>
From: "Alex" 
Sent by: cp... at googlegroups.com
Date: 06/26/2020 10:29PM
Subject: [CP2K:13553] ssmp and CPU loads

Hi all,

We've just built the latest version according to https://www.cp2k.org/howto:compile?s[]=ssmp

The main use case is single-node multi-CPU (36 cores /72 threads available), so SSMP appeared like the correct option, i.e., none of the run commands involving mpirun. Our understanding is that the executable will use all available threads. We have confirmed it using a small system (~50 atoms) -- the system load shows one instance of cp2k with something like 2400% load, which is similar to for example how Gromacs runs using OpenMP. So, all is great.

However, when running a large system containing ~1000 atoms, there is a single cp2k instance with 100% load and ~1TB RAM usage. SCF updates are sluggish, but it's a large system, so we can't tell if it's due to size or the fact that the CPU load is improperly low. Here is my question: is this correct behavior? Is there a way or need to force a particular number of threads to run on? Are we building the correct version for our needs?

Another, more minor question: does multithreading help CP2K performance? The same machine is used for coupled cluster calculations and the users insist that HT hurts their performance, so it's currently off. 

Many thanks,

Alex  
  -- 
 You received this message because you are subscribed to the Google Groups "cp2k" group.
 To unsubscribe from this group and stop receiving emails from it, send an email to cp... at googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/dad49414-499b-4804-afc0-fc4855d0772bo%40googlegroups.com.