HEP-SCORE deployment TF meeting

Wednesday 18 Jan 2023, 16:00 → 17:00 Europe/Zurich

Zoomland

Domenico Giordano (CERN), Randall Sobie (University of Victoria (CA)), Randy Sobie (University of Victoria (CA))

2023-01-18_HSTF-draft-notes.pdf

16:00 → 16:05 Welcome, note-taking, notes from previous meeting

Minutes of 2022-12-07

CHEP2023 abstract accepted as oral contribution

 

"

We're pleased to announce that your abstract "HEPscore: a new benchmark for WLCG compute resources" with ID #120 has been accepted in track "Track 7 - Facilities and Virtualization" (Oral).
Conference: 26TH INTERNATIONAL CONFERENCE ON COMPUTING IN HIGH ENERGY & NUCLEAR PHYSICS (CHEP2023)
Submitted by: Domenico Giordano
Track classification: Track 7 - Facilities and Virtualization
Presentation type: Oral

"

 

Candidatures for the speaker are open. Thanks to send proposals to Randy and Domenico

16:05 → 16:20 Summary of the Dec WLCG MB report

Speaker: Domenico Giordano (CERN)

Domenico's slides

Plans for HEPScore were presented to the WLCG MG on 2022 Dec 20.

https://wlcg-docs.web.cern.ch/boards/MB/Minutes/2022/MB-Minutes-20221220-2.pdf

 

Here is reported the contribution summary

"

Domenico Giordano summarises the plans for the adoption of the new HEPScore benchmark in WLCG.
The workloads considered for the first official version (HS23) are from ATLAS, ALICE, CMS, LHCb and
Belle II and the score will be normalised to be the same as HS06 on the reference server, to make the
transition easier for the accounting system. The supported architectures are x86 and ARM and the score
has been shown to be reproducible at the per-mille level. The goal is to have HS23 in production by 1 April.
ARM support will be included in the first version only if all the workloads will be ready by 14 February. Sites
will not be required to re-benchmark already deployed hardware, but they may choose to do it.
The following comments are noted:
• James Letts asks for more details about the procedure to update a workload. Domenico explains
that a change should be requested by the experiment contact in the HEPScore Task Force and
Simone adds that updates in HEPScore should not happen too frequently, to be manageable for
the infrastructure. Simone proposes that, whenever an experiment believes that an update to
HEPScore is needed, they should submit the request to the MB, which will then task the relevant
experts with understanding costs and benefits. Simone’s proposal is accepted by the MB.
• Elizabeth Sexton-Kennedy asks if HEPScore might be used also to pledge fractions of HPC
facilities using ARM rather than only for traditional Tier-1/2 sites. Simone replies that today we do
not have a model to pledge a fraction of an HPC, but this might happen in the future and HEPScore
should then be re-evaluated to accommodate this case.
The Management Board gives its green light for the plan outlined by Domenico

"

16:20 → 16:35 HEPscore test configurations

Speaker: Domenico Giordano (CERN)

WLCG-TF-18-01-2023-giordano.pdf

16:35 → 16:55 HEP-Workloads status

Speakers: Andrea Valassi (CERN), Randall Sobie (University of Victoria (CA)), Stefano Piano (INFN (IT))

20230123-BMKTF-LHCbSim-AV-v001.pdf

20230123-BMKTF-LHCbSim-AV-v001.pptx

Sobie-B2-ARM-Workload.pdf

ALICE workload status:

• Ported the code to ARM in December, ran into some issues which our experts fixed in a short time. 
  • The code is on CVMFS and since last week we have stable setup in multi arch configuration. 
  • CI produced new singularity / apptainer images for both x86 and arm that run fine in our test machine. 
• The load has been reduced to a maximum of 4 for our workload in order to meet the running condition of the benchmark which are different wrt the grid jobs. We have found no stability problems with such a load on our test machine. Needed to measure stability on machines with high number of cores.
• For what concerns the reco-only workflow instead of digi-reco, this does not bring additional benefits and it is less representative, so we decided to stay with digi reco workflow.
• What we could still aim to do is to provide a specific benchmark that runs on both CPU and GPU (TPC only reconstruction). This will require some effort of our GPU reconstruction experts to provide this.

16:55 → 17:00 Any other business