Draft Minutes NEMO HPC WG  Friday 20 May 2016

Attending:  Lucien Anton (Cray), Jeremy Appleyard (Invidia), Mike Bell (Met Office, chair), Miguel Castrillo (BSC), Mondher Chekki (Mercator), Maff Glover (Met Office), Tim Graham (Met Office), Dmitry Kuts (Intel), Claire Levy (CNRS), Yann Meurdesoif (CEA), Silvia Mocaverro (CMCC), Andy Porter (STFC), Stan Posey (Invidia), Martin Schreiber (Uni Exeter), Julien le Sommer (CNRS)  


1.    Status of actions from previous meetings

Mike has started to put minutes and other information on the NEMO HPC WG page. 
Mondher has given information on computers accessible to Mercator. 

2.    NVIDIA presentation (30 mins including discussion)

Jeremy Appleyard presented. It was noted that the SOR solver is not included in the vn3.6 and future versions of NEMO. The issue of whether OpenACC directives should be moved up to vn3.6 or included in the trunk was raised and briefly discussed. OpenMP directives are currently being ported into the trunk. 

3.    Progress setting up standard configurations for performance testing (Silvia and Miguel)

Benchmark configurations for ORCA025 LIM3 and GYRE will be ready to use after Tim has done some final checking of GYRE.

4.    Progress on single core performance testing (Martin and Silvia)

Martin has proposed a revision to the timing interface using integer IDs instead of strings for code regions with an additional params argument to enable nested timing. The new version could be introduced using the grep pattern matching Unix command. 
Action: Martin, Tim, Silvia, Claire, Miguel and Andy Porter to agree details of the timing interface.    

5.    Memory leaks at NEMO version 3.6 (Silvia)

Large NEMO configurations at vn3.6stable run for a long period of time sometimes suffer from memory leaks which cause crashes running out of memory. The NEMO system team is looking for ideas from the HPC team on how to tackle the problem. It is not known whether the problem is specific to certain keys or compiler options and it may not be completely reproducible.  It was suggested that tools such as valgrind and memory checker could help pin down the problem.     

Action: Tim and Silvia to gather (some) information on when the problem has occurred to start a discussion at the next meeting.   

6.    Status of proposals for funding (e.g. Copper proposal)

A decision on the Copper proposal is expected in August. Andy Porter and Rupert Ford have proposed 6 months work within the ISENES-2 project to re-structure key routines in the full NEMO code such that they are suitable for use with the PSyclone code-generation system.

7.    What topics should be discussed in future meetings ?

OpenMP directives, single core performance testing and memory leaks will be discussed in future meetings 

8.    Date for next meeting

Mike will send out a doodle poll looking for dates in September. 

Last modified 4 years ago Last modified on 2016-09-28T11:10:07+02:00