Software Review January 13, 2015

From GlueXWiki
Jump to: navigation, search

GlueX Software Review Meeting
Tuesday, January 13, 2015
2:00 pm EST
JLab: Mark's Office

Remote Connection

Agenda

  1. Data Challenge 3
    • Farm Information
      • Just follow up on this, the new farm nodes report 48 cores, but according to Chip, they have dual 12 core CPUs and therefore 24 real cores each and 24 hyper threads. A few lines from /proc/cpuinfo are shown below indicating the exact processor type. They are currently set to each have 42 job slots total.

        From an e-mail exchange I had with Jie Chen at the beginning of Dec. I’m reminded they have an “exclusive” tag concept in the farm. I think only nodes having this tag are allowed to give all of their slots to a single job. This tag has been set for all of the new farm nodes. If your request for number of cores is equal to the number of slots, then you are given exclusive use of the node.

        I asked about the seemingly odd choice of 42 slots for a node and Chip explained it this way:

        We have empirically observed that we maximize throughput by using 7/8 of all of the hyperthread slots. Going higher makes file I/O head thrashing worse but gains no additional CPU performance on a mix of serial jobs. 7/8 is marginally better than 3/4 which is decidedly better than 5/8.

        The 42 seems like they actually decided to go with the 3/4 instead of 7/8, but the reasoning is the same.

        So, we can set jobs to request 42 cores, but then run them with “-PNTHREADS=Ncores” to automatically launch 48 threads. This should give us the maximum use of all cores and hyper threads on the node. If we specify a smaller number of cores,then we should probably only launch that many threads in order to be good neighbors with other jobs sharing the node.

        processor : 47
        vendor_id : GenuineIntel
        cpu family : 6
        model : 63
        model name : Intel(R) Xeon(R) CPU E5-2670 v3 @ 2.30GHz
        stepping : 2
        cpu MHz : 2300.000
        cache size : 30720 KB
        physical id : 1
        siblings : 24
        core id : 13
        cpu cores : 12
  2. 2015 Software Review
    • Dates February 10+11, 2015.
      • Dear all,
        we still seem set on the February 10 + 11 dates for a next software and computing review, even if we may loose one reviewer on that date. Chip is confirming with the reviewers now.

        What I would like is that you think about the possible presentations, assuming that the structure of the review is similar as last time around. Then we should plan on picking an hour or so for a meeting next week to bounce the proposed agendas of each other, and discuss any questions on what we should emphasize following the previous review report.

        So, please prepare a draft agenda this week! Assume half an hour plenary for Halls A & C, half an hour plenary each for Halls B and D, and then short breakouts, etc.

        I will set up a meeting for next week with call-in, and then we should plan on one dry run only to see each other's near-final talks and provide feedback, perhaps January 22nd or 29th (we have to evade the EIC cost review) or so. For now you should assume February 4 as date to submit the presentations for final feedback to the review webpages (it will be indico again), and then we can open it to the reviewers shortly after.

        This closely follows our preparations of last review, which were more than sufficient in my view and did not take much of your times - you should know what to do. I will pass by a few people later this week to collect some views on what dates/times would work.

        Best regards, Rolf
    • Topics for the 2015 Software Review