Choose timezone
Your profile timezone:
Powered by Indico
v3.3.3-pre
Meeting Date: 10/03/2017
Attendees: D. Abdurachmanov, M. Alef, Aritz, L. Atzori, J-M Barbet, D. Giordano, C. Grigoras, M. Guerri, D. Lange, M. Michelotto, A. Sciaba
*****************************
Intro:
*****************************
Domenico mentions the twiki of the WG with new material to document the activity and results. In particular the “Action” session lists the actions to be taken in the following weeks.
Domenico proposes to re-establish the approach followed to study HS06, i.e.:
- identify the reference workloads of each experiment
- select a defined set of resources to run benchmarks under control, avoiding the uncertainties of the grid job submission.
Having this approach already established will be beneficial not only to test the fast benchmarks but also to be ready for the evaluation of the successor of HS06 (new SPEC release should be available in 2017)
David A. mentions that for CMS a mini-framework has been in place in the context of OpenLab student activity and could be used to establish (or at least inspire) a general approach for all the experiments.
David will circulate details.
Urgent:
Need to evaluate what can be studied before the April GDB (12 of April)
⁃ Manpower from the experiments
⁃ Resources available
⁃ Studies to be performed
News: The HEPiX Spring 2017 Workshop (24-28 April 2017, Budapest
Hungary) will include a benchmarking session.
*****************************
Node configuration at GridKA:
*****************************
Manfred mentions that the performance of DB12-at-boot on the GridKA nodes is still not understood (see plot in slide 8 of talk[1])
In particular the discrepancy running on 20, 32 or 40 parallel benchmark copies is not understood.
In addition the WNs configured with the 1, 1.5, 1.6, 2 job slots per physical core appear to have a poor performance. And this requires further investigation
- Clarification asked about the measurement in plot at [1]:
- is SMT always ON, or just when the number of parallel processes is greater then the number of physical threads?
Answer: Yes, SMT was by default enabled in the BIOS.
- Could also the DB12 measurement be affected by the WN poor performance?
Answer: Yes, DB12-in-job score is by a factor of 5...10 lower than expected
*****************************
Update on Profiling ATLAS Kit
Validation on Intel Haswell-EP
processors
*****************************
Marco G. shares an initial draft of the Internal Note that he is writing to summarise his findings on the study of the thread performance in Haswell (and Broadwell) processors.
Marco will leave soon CERN (in 1 week time): people interested in providing feedback to this draft document are invited to do it in the coming days.
The Internal Note will be published in CDS. The results will be presented also at the next HEPiX Workshop
*****************************
Passive benchmarking of
ATLAS Tier-0 CPUs
*****************************
Andrea S. reports about his study on passive benchmarking.
The work was already done for CMS and Atlas jobs running on the grid.
The focus of the current report is the Atlas Tier0, running reconstruction jobs.
The Atlas Tier-0 configuration changed in March ’16 from a configuration with SMT enabled to a configuration with SMT disabled. Choice driven by job memory requirements and consequent job efficiency improvement.
The results of the passive benchmark across different CPU models is then correlated with the HS06 values provided by the CERN procurement DB.
The scaling among passive benchmark and HS06 looks generally good with two exceptions: Opteron is way off, and Haswell tends to perform better than Sandy Bridge, with a performance gap of ~10% both in configurations with SMT ON and OFF
Michele M. asks if HS06 values with -m64 are available for this HW.
This information is not in the CERN procurement DB, but Domenico is running HS06 with -m32 and -m64 on few resources to verify the difference in performance.
Domenico asks if the same study can be repeated for the other experiments, for what concerns the Tier-0.
Andrea will include in the twiki the information needed from the experiment job monitoring to run the passive benchmark. For instance it seems that for CMS it’s not possible to run the analysis because the CPU model name is missing in the job monitoring report.
*****************************
AOB
*****************************
Next meeting: 24/03/2017
*****************************
Actions
*****************************
- For the site and experiments’ representatives: to fill the information in the Twiki.
- David A.: to circulate details about the tool implemented by the OpenLab student
- Andrea: to document in the twiki the information needed to run the passive benchmark.
[1] https://indico.cern.ch/event/578967/contributions/2455893/attachments/1407381/2151099/pre-gdb-2017-02-07.pdf