4 Replies Latest reply: Jan 28, 2013 3:11 PM by user738616 RSS

    Memory on the active site of push replication

    965779
      In the case of active/passive push replication configuration, should one budget more memory for the active cluster?
        • 1. Re: Memory on the active site of push replication
          user738616
          Hi,

          No really as the objects are not de-serialized unless you are planning to do some kind of transformation. But definitely, you would require more processing capacity on Active side.

          HTH

          Cheers,
          _NJ                                                                                                                                                                                                                                                                                                                                                                                                                                   
          • 2. Re: Memory on the active site of push replication
            965779
            What if the passive site goes down? I assume the active site will queue up the outstanding messages. In such a case, one need more memory on the active site. Correct? Are there any guidelines on how much additional memory one should budget in such a scenario. I suppose this number would be a function of the data set size.

            Edited by: 962776 on Jan 25, 2013 4:22 PM
            • 3. Re: Memory on the active site of push replication
              DJW
              As with any heap estimation question, I'd suggest you run tests and perform measurements, there are too many variables to reliably estimate from first principles. Something like:

              1. Populate a storage node with replication active
              2. Allow replication to catch up (queues empty)
              3. Break replication
              4. Force a full GC
              5. Measure heap
              6. Perform a known number of updates to existing cache entries (so as not to significantly change heap used by cache)
              7. Force a full GC
              8. Measure heap

              Run at least two storage nodes so that the impact of backups and indexes is correctly represented
              Differences in heap between steps 5 and 8 will give you a heuristic for heap size per queued entry.
              These kind of measurements can vary significantly according to Coherence version, JVM version, GC configuration, etc.

              You'll want to ensure you push enough updates so that the change in heap use is a significant compared to heap use before, you'll probably need to perform a number of runs.
              • 4. Re: Memory on the active site of push replication
                user738616
                962776 wrote:
                What if the passive site goes down? I assume the active site will queue up the outstanding messages. In such a case, one need more memory on the active site. Correct? Are there any guidelines on how much additional memory one should budget in such a scenario. I suppose this number would be a function of the data set size.

                Edited by: 962776 on Jan 25, 2013 4:22 PM
                Hi,

                If you mean that the entire Coherence cluster goes down on passive site then anyways, you have to go back to DB to load the data into the cache because PushReplication will only replay the events that have happened into the ActiveSite and not replicate the delta of the data.

                If you mean the link failure between Active/Passive cluster then typically extra memory would depend on how much events and data are you expecting during the passive site failure or would depend on SLA provided by networking team.

                HTH

                Cheers,
                _NJ