Floating Point Speed calculation


Advanced search

Message boards : Cruncher's Corner : Floating Point Speed calculation

AuthorMessage
Profile tolafoph
Send message
Joined: Sep 14 07
Posts: 110
Credit: 20,010,921
RAC: 0
Message 114514 - Posted 15 Oct 2011 7:27:49 UTC

    Hi,

    the Floating point speed on the server status page shows
    529.1 TFLOPS
    Are these the total FLOPS for CPU+GPU or just the CPUs?
    Only the CPU TFLOPS from the three apps add up to 113+176+239=528.
    The total Floating point speed has the footnote

    (from recent average credit of all users).

    But this can`t obviously be the case. The BRP4 has less then half the FLOPS of FGRP1 but twice the Credits.
    I think for BRP4 when it was GPU-only a few weeks ago the CPU TFLOPS were < 10 but the Granted credit/100
    was around 1 Million.

    For FGRP1 and S6Bucket which are only CPU apps I get an ratio of 330000 credits/(TFLOPS * week) or 47000 credits/(TFLOPS * day).

    If I use this ratio to calculate the TFLOPS for BRP4 with the Granted credit/100 = 1,603,745; I get 486 TFLOPS. This is the number equivalent to CPU TFLOPS.

    All apps would add up to 486+176+239=901 TFLOPS.

    So the project has way more calculation power then it shows on the server status page.

    p.s. I noticed that BOINC-Stats only shows 240 TFLOPS.


    Sascha

    Profile Bernd Machenschalk
    Forum moderator
    Project administrator
    Project developer
    Avatar
    Send message
    Joined: Oct 15 04
    Posts: 3241
    Credit: 90,383,219
    RAC: 15,651
    Message 114543 - Posted 17 Oct 2011 13:13:47 UTC

      Last modified: 17 Oct 2011 13:14:02 UTC

      Your assumptions are mostly correct, with one exception: The "Floating point speed (from recent average credit of all users)" is indeed what it says. This is independent of the way (i.e. the WUs) this credit was gained from.

      The "CPU TFLOPS (from successful tasks last week)" are actually "normalized" such that these add up to the "Floating point speed". It is true that this currently counts CPU FLOPS only, so it largely underestimates the FLOPS of BRP4. But it's the BRP4 share of the total "Floating point speed" that's underestimated, not the total computing power.

      BTW: does somebody happen to know how BOINCStats calculates the computing power of a project?

      BM

      Akos Fekete
      Volunteer developer
      Avatar
      Send message
      Joined: Nov 13 05
      Posts: 562
      Credit: 4,404,768
      RAC: 0
      Message 114551 - Posted 17 Oct 2011 18:02:10 UTC - in response to Message 114543.

        Last modified: 17 Oct 2011 18:03:43 UTC

        BTW: does somebody happen to know how BOINCStats calculates the computing power of a project?

        I think it uses a very simple method.
        BOINCStats always queries the data of each computers.
        It simply adds the floating point speed of the active computers.
        ____________

        Profile hoarfrost
        Send message
        Joined: Feb 9 05
        Posts: 187
        Credit: 26,744,601
        RAC: 9,750
        Message 114555 - Posted 17 Oct 2011 19:42:02 UTC - in response to Message 114543.

          BTW: does somebody happen to know how BOINCStats calculates the computing power of a project?

          BOINC Stats FLOPS = RAC / 2.
          But, if I understand right, it is a mistake - because 100 CS = 1 day * (1 GFLOPS + 1 GINTOPS) and if we have 50 000 000 CS/day performance, we have a 500 TFLOPS and 500 TINTOPS. But not "reverse" when 50 000 000 CS/day = 500 TFLOPS or 500 TINTOPS.

          I think that in code of BOINC Stats mistake is present.
          ____________

          Post to thread

          Message boards : Cruncher's Corner : Floating Point Speed calculation


          Home · Your account · Message boards

          This material is based upon work supported by the National Science Foundation (NSF) under Grants PHY-1104902, PHY-1104617 and PHY-1105572 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

          Copyright © 2014 Bruce Allen