Go to the Indico Home Page Oldest event Older event Up to category Newer event Newest event Printable version
  • Compact style
  • Indico style
  • Indico style - inline minutes
  • Indico style - numbered
  • Indico style - numbered + minutes
  • Indico Weeks View
    • Switch to the management area of this event
    Login

    HEP-Workloads mid-cycle Sprint meeting

    → Europe/Zurich
    Manfred Alef (Karlsruhe Institute of Technology (KIT)), Domenico Giordano (CERN), Michele Michelotto
    Description

    Active Sprint: https://its.cern.ch/jira/secure/RapidBoard.jspa?rapidView=6519

      • 14:30 → 15:00
        Sprint mid-cycle meeting 30m
        Speakers: Domenico Giordano (CERN), Manfred Alef (Karlsruhe Institute of Technology (KIT)), Michele Michelotto

        Running loop of CMS workloads processing 1000 events each (default: 100 events) on host c01-010-156.gridka.de. This host has plenty of memory (128 GB = 3 GB per task) and free local disk space. Wallclock time is around 13 h, scores schould not be affected by initialization or tails.

        9 runs completed so far, with 4 outliers. Spread of the 5 perfect jobs is around 1%. Worst one (3 short tasks (crashed?)):

        test_1553121762_5143:
        -rw-r--r-- 1 alef scc 3.2G Mar 21 09:09 test_1553121762_5143/proc_10/step1.root
        -rw-r--r-- 1 alef scc 3.2G Mar 21 09:10 test_1553121762_5143/proc_1/step1.root
        -rw-r--r-- 1 alef scc 3.2G Mar 21 09:19 test_1553121762_5143/proc_2/step1.root
        -rw-r--r-- 1 alef scc 3.2G Mar 21 08:53 test_1553121762_5143/proc_3/step1.root
        -rw-r--r-- 1 alef scc 2.6G Mar 21 07:32 test_1553121762_5143/proc_4/step1.root
        -rw-r--r-- 1 alef scc 3.2G Mar 21 08:55 test_1553121762_5143/proc_5/step1.root
        -rw-r--r-- 1 alef scc 3.2G Mar 21 08:55 test_1553121762_5143/proc_6/step1.root
        -rw-r--r-- 1 alef scc 119M Mar 21 00:07 test_1553121762_5143/proc_7/step1.root
        -rw-r--r-- 1 alef scc 1.7G Mar 21 05:10 test_1553121762_5143/proc_8/step1.root
        -rw-r--r-- 1 alef scc 3.2G Mar 21 09:12 test_1553121762_5143/proc_9/step1.root

        Resulting scores:

        • througput_score 1.1501 (+15%)
        • CPU_score 0.9802 (x4)
          • Reason: the shortest task (proc_7) has processed only 144 (instead of 4*1000) events. To calculate the ev/s score, parseResults.sh divides the input numbers of ntasks*nevt by the runtime (CPU Summary:Total loop) of the task; suggestion (workaround): grep also 'Number of Events' from the logfile instead of using the static input values, and add an appropriate error flag to the JSON snippet.