Extremely short work units


Advanced search

Message boards : SZTAKI Desktop Grid : Extremely short work units

AuthorMessage
verdy_p
Send message
Joined: Feb 9 06
Posts: 5
Credit: 9,657
RAC: 0
Message 2137 - Posted 6 Mar 2006 13:58:58 UTC

    Last modified: 6 Mar 2006 14:07:12 UTC

    I\'m not surprized that the project is most often depleted from workunits now.

    These work units are SOOOOO short that it takes often just one or a few dozens of SECONDS to complete (so with very small credits per work unit).

    Couldn\'t work units be at least 16 times longer? Even so, it would be difficult to participate if they are completed after this time,because it would require a constant Internet connection, that would take significant bandwidth to constantly exchange workunits and results. Even your bandwidth may be affected by so many queries coming everytime from lots of hosts.

    Well, my PC have been idle for about one week without any workunit to complete because of this; today, I could finally get a few work units, but I had got only 1 per PC instead of about 30, which would be accurate given that this is roughtly the number of WUs my PC would complete in 1 hour, as indicated in the request.

    Unfortunately, the server apparently computes the numberof work units to give to each PC according to its average results during a period where the PC was left idle for more than 99% without any work unit to compute.So the server just gives 1 work unit, that will be completed and its results sent back in the next minute, and then no more work units.

    Then the PC will repeatedly query the server for new work, with exponential delay between queries. Given that there\'s a nearly 80% chance that the server no longer haswork units in the next minute, the PC will wait longer, and then the chance of getting new work units always decreases, and this enters an infinite loop where every PC will be stalled with wait delays between queries reaching several days. If there are somework units available, it will not see it.

    The finalresult is that the project is not scheduled at all towork correctly and automatically with BOINC, and it unnecessarily wastes network bandwidth (both our own, and yours...).

    Work units NEED to be much longer (by at least 1 or 2 orders of magnitude) than what they are now.

    For example, look at the results on one of my PC:
    http://szdg.lpds.sztaki.hu/szdg/results.php?hostid=12052

    The average computing time is MUCH too small, and the server really does not give it enough work.
    ____________



    Profile kadam
    Project administrator
    Avatar
    Send message
    Joined: May 25 05
    Posts: 589
    Credit: 38,614
    RAC: 0
    Message 2138 - Posted 6 Mar 2006 14:29:36 UTC - in response to Message 2137.

      The WUs of last week were only 4 times bigger than the ones right now...and what did users do?...complained about the big size of the WUs. I\'m really trying to find the optimal size, but trust me it\'s not easy...
      ____________
      If you like BOINC, you may also find CaretCursor to be appealing.

      Profile Bob Carlton
      Avatar
      Send message
      Joined: Jan 13 06
      Posts: 18
      Credit: 9,224
      RAC: 0
      Message 2142 - Posted 6 Mar 2006 17:11:02 UTC - in response to Message 2138.

        The WUs of last week were only 4 times bigger than the ones right now...and what did users do?...complained about the big size of the WUs. I\'m really trying to find the optimal size, but trust me it\'s not easy...



        Adam, we know that the change to dimension 12 gave you a lot of headaches, and we are bearing with you. The only complaint I have is the lack of flow in the workunits. At times, I can do 100WUs in a heartbeat then no work for 12 hours. Just a better flow of the work would be great with me. The small size of the workunits does not matter, just keep them coming.
        ____________
        member of

        Profile UBT - Halifax--lad
        Avatar
        Send message
        Joined: Sep 10 05
        Posts: 126
        Credit: 3,147
        RAC: 0
        Message 2144 - Posted 6 Mar 2006 19:13:31 UTC - in response to Message 2138.

          The WUs of last week were only 4 times bigger than the ones right now...and what did users do?...complained about the big size of the WUs. I\'m really trying to find the optimal size, but trust me it\'s not easy...


          How can you say the WU\'s were 4 times bigger last week that would make them all last 16 minutes on my computer rather that 4 mins now, a nice length would be a couple of hrs or is it possible to do what Rosetta do and tell BOINC how long a WU should run for?

          ____________
          Join us in Chat (see the forum) Click the Sig


          Join UBT

          Profile Ananas
          Send message
          Joined: Jul 12 05
          Posts: 222
          Credit: 665,833
          RAC: 0
          Message 2145 - Posted 6 Mar 2006 20:37:57 UTC

            Last modified: 6 Mar 2006 20:46:53 UTC

            The problem was not so much the size of the WUs (although 15 hours is not excactly four times 10 seconds), it was more the too short deadline and the wrong estimated time. It delivered 20 WUs where it should have delivered 1 WU - and even one WU of 15 hours is hard finish successfully in 3 days if

            - the project shares its hosts with other projects. I had SZTAKI running as a backup project with a not too high ressource share but it was running on 7 PCs / 12 CPUs so it did get its CPU cycles

            - the results are not restartable - I could have rescued a few results by increasing the <debt> values but experience told me that a restart kills them.

            With a fair deadline, correct estimated time and restartable results, not many would have complained I guess (besides the CPU time jumping to 0 after the long calculation often).


            Fix the restart problem then bring those long ones online but configure them right and everything will be much better.

            If a result might run between 10 and 20 hours, better make it estimate 20 rather than 10 or we will start to panic again ;-)

            Profile UBT - Halifax--lad
            Avatar
            Send message
            Joined: Sep 10 05
            Posts: 126
            Credit: 3,147
            RAC: 0
            Message 2146 - Posted 6 Mar 2006 21:36:31 UTC

              All we need is a 10 day deadline to match what you can set your cache to and then some WU\'s that are around 2 hrs or so, that way they won\'t run out as quick as the short 4 min ones I got
              ____________
              Join us in Chat (see the forum) Click the Sig


              Join UBT

              Post to thread

              Message boards : SZTAKI Desktop Grid : Extremely short work units


              Home | My Account | Message Boards


              Copyright © 2017 SZTAKI Desktop Grid