2 Replies Latest reply on Jul 8, 2011 3:36 AM by Ron K

    Failover measurement

    Ron K Newbie

      Hi all,

       

      i was just measuring how long how long a failover takes and measured it by using an application once with a SFSB with replication and once with a SLSB. These beans are deployed on the server. Additionally to this, theres a client application, that looks up teh bean, and starts threads to count numbers on serverside. Whenever a number is received at a node on the server side, the node writes it into the database (timestamp is made by teh database to avoid impreciness of the time). 

       

      My first test case was to start up 2-5 nodes and force them to failover and retrieve the deltas between 2 numbers in a row. The result was that a usual count tooks about 0.125s and whenever a failover happened, the duration was about 0,7s.

      That worked well eventhoough according to different load-balancing strategies.

       

       

      But in another testcase, started a cluster 2-5 nodes and according to this cluster i tested every clustersize with 1-40 concurrent counting threads.

      I cant explain myself the result, because the failover times were shrinking in relation to the amount of threads. As an example:

       

      Threads

      1FirstAvailable0,750
      2FirstAvailable0,688
      3FirstAvailable0,656
      4FirstAvailable0,672
      5FirstAvailable0,625
      6FirstAvailable0,641
      7FirstAvailable0,578
      8FirstAvailable0,656
      9FirstAvailable0,594
      10FirstAvailable0,593
      15FirstAvailable0,578
      20FirstAvailable0,469
      40FirstAvailable0,188
      Threads
      1FirstIdentical0,672
      2FirstIdentical0,640
      3FirstIdentical0,656
      4FirstIdentical0,625
      5FirstIdentical0,641
      6FirstIdentical0,704
      7FirstIdentical0,641
      8FirstIdentical0,641
      9FirstIdentical0,609
      10FirstIdentical0,609
      15FirstIdentical0,562
      20FirstIdentical0,500
      40FirstIdentical0,203

       

      As smbdy can see, the time needed to accomplish a failover shrinks by teh numbers of threads. Is that an issue of jgroups or of the load-balacning proxy?

       

       

      Thanks in advance,

        • 1. Re: Failover measurement
          Ron K Newbie

          Couldn't adjust it anymore, but the results above are yet only for the SLSB failover!

           

          And this ones for SFSB:

          234
          1FirstAvailable0,7030,8750,781
          2FirstAvailable0,7660,8120,734
          3FirstAvailable0,7500,7820,781
          4FirstAvailable0,7030,7650,781
          5FirstAvailable0,7340,7190,765
          6FirstAvailable0,6410,6880,719
          7FirstAvailable0,6880,6710,781
          8FirstAvailable0,6410,6570,797
          9FirstAvailable0,6720,7660,765
          10FirstAvailable0,5620,6710,719
          15FirstAvailable0,5940,6250,672
          20FirstAvailable0,5940,6090,656
          40FirstAvailable0,5940,6090,594

           

          Doesnt happen here, only for SLSB:

          234
          1FirstAvailable0,7500,7810,691
          2FirstAvailable0,6880,6560,687
          3FirstAvailable0,6560,7030,656
          4FirstAvailable0,6720,6410,650
          5FirstAvailable0,6250,6560,641
          6FirstAvailable0,6410,6410,637
          7FirstAvailable0,5780,6250,656
          8FirstAvailable0,6560,6400,641
          9FirstAvailable0,5940,6090,625
          10FirstAvailable0,5930,6400,640
          15FirstAvailable0,5780,5630,578
          20FirstAvailable0,4690,5130,500
          40FirstAvailable0,1880,1720,172
          • 2. Re: Failover measurement
            Ron K Newbie

            Ok, Measurement mistake :-) solved