Version 1 (modified by mocavero, 3 years ago) (diff)

NEMO HPC subgroup: Thurs 13 Oct 2016

Attending: Mike Bell (Met Office), Miguel Castrillo (BSC), Tim Graham (Met Office), Silvia Mocavero (CMCC), Andy Porter (STFC), Martin Schreiber (Uni Exeter)

1. Single-node performance

1.1 FLOPS over-counting issue & systems (including KNL) access

To solve the over-counting issue, the idea (proposed by Andy and agreed by the subgroup) is to combine the results provided by a fortran parser able to provide FLOPS count for each loop and the measurement provided by the perf_regions tool for the other metrics

Action: Andy to test the parser on the NEMO code. First results could be available next week.

The group has now access to several Intel architectures, except for KNL. Silvia checked the possibility to access KNL nodes provided by Intel. The access requires to sign an NDA. The new Archer to be delivered this month will include some KNL processors, then the group decided to not follow up on the Intel NDA offer

Action: all to check the possibility to access other systems (e.g. ARM), also including KNL. Martin already provided info on the ARM system.

1.2 status of perf_regions integration in NEMO

The activity is almost completed and the tool has been tested on the GYRE benchmark configuration. Some minor changes to the python script are missing and the integration of the ARCH files for the combined NEMO and tool compilation is planned.

Action: Tim and Martin to finalize the python script. All to add the ARCH files for the systems they have access. Some initial assessments of the GYRE performance should be done before the Developers' committee meeting (Tim with some assistance from Martin and Andy). This will be quite challenging given that the Developers' meeting is early December but it's a good target to aim for.

1.3 perf_regions document

Martin sent the document to the subgroup. The document contains a brief description of the tool.

Action: all to check the document to add missing information.

Martin proposed to start thinking at a publication on the perf_regions tool, by considering two case studies (NEMO and an astrophysics code that will use the tool).

Action: Martin to send an outline of the paper to understand its scope, all to evaluate the interest to contribute.

2. HPC chapter of the NEMO development strategy document - writing process

A draft of the document strategy will be discussed during the next Merge Party (30 nov - 2 dec) and the Developer's Committee (7-8 dec). The chapter should include the list of key points for the next 10 years HPC development strategy. Some attention has to be paid to efficiency, flexibility and scalability. The main goal is to run efficiently the code with a local domain 10 by 10. The current limit is ~40 by 40. The first draft of the HPC chapter has been written by Mike and revised by Silvia and Miguel.

Action: Mike to integrate Miguel's revisions and to circulate the document within the sub-group. All to check and agree on the list of key points.

3. NEMO Workplan actions (2016-2017)

The definition of the 2017 NEMO workplan is under discussion. The HPC-WG could contribute to the next workplan with two new actions: (i) analysis of single-core performance by using the perf_regions tool (PI: HPC-subgroup) and (ii) replacement of wrk_alloc with allocate: test of performance improvement and extension strategy (PI: Tim and Silvia).

Action: Tim and Silvia to propose the new actions to the System Team during the next Merge Party.

9. Date for next meeting

Next meeting will be in the second half of November.

Action: Silvia to send the doodle poll for the next meeting.