Benchmark WG Meeting

Thursday 4 Sept 2008, 17:00 → 18:00 Europe/Zurich

31-1-030 (CERN)

Helge Meinhard (CERN-IT)

To join both the web and audio conference (recommended), click here: https://audioconf.cern.ch/call/0181047 If you only want to participate in the audio conference, you may dial in with the following numbers: +41227676000 (Main), and enter access code: 0181047

CPU Benchmark WG Meeting Notes, September 4, 2008

Present:
Alex Iribarren, Helge Meinhard, CERN/IT
Franco Brasolin, CERN/ATLAS
Peter Wegner, DESY Zeuthen
Michele Michelotto, INFN Padova
Franco Brasolin, CERN/ATLAS
Martin Bly, RAL
Ian Gable, U Victoria

1. Helge's review of events in June

• Helge received pressure to produce a finding in June by the GDB and WLCG Management Board.
• He reported that the detailed statistical analysis could not be delivered in the timescale requested.
• The all_cpp benchmark was put forward by Helge as it delivered a ratio of floating point and integer operations consistent with those measured my Andreas Hirstius on lxplus using perfmon.
• Helge was asked if he would like to chair a GDB Working group on benchmarking. He declined but offered the input of the HEPiX Working group.

2. Discussion of all_cpp benchmark.

• Michele commented all_cpp is hard to explain to vendors, and there may not be future versions of it  available. However he does recognize the benifits of the all_cpp.
• There were comments that because the all_cpp benchmark is composed of 3 int and 4 fp test that some sacling factor may need to be applied. Helge comented that he 'naively' believes that because the composition of actual operations is the same as thos measured on lxplus (i.e. 10% fp) that no scaling is required.
• Michele raised the question of how the results are being reported. Michele would like to see the results reported as the mean (geometric, or arithmetic) of the fp and int components of the all_cpp benchmark rather then as a sum of the two as Alex's script currently produces.
  

3. Action items

• Michele will run the all_cpp benchmark on his 2.5 GHz Barcelona
• Martin will organize his colegues to run all_cpp on his 2.3 GHz Barcelona
• Peter Wegner will also run the all_cpp on his two Barcelonas
• Michele will produce a draft abstract for CHEP 2007
• Helge would prefer that results are delivered to him in advance of the next meeting (september 15) such that he can perform an analysis prior to the meeting.
• Resolve if we should report the mean of the fp and int results for all_cpp or simply sum. Resolve during the next meeting?

4. Next Meeting

• The next meeting will be held 15:00 UTC September 15.