2 Replies Latest reply on Dec 19, 2005 12:07 PM by Tim Fox

    Multiplex transport performance.

    Tim Fox Master

      Hi guys

      The multiplex docs state that the performance of the multiplex transport is between about 2 and 9 times "slower" than normal sockets depending on what you're doing.

      Does "slowness" here refer to a reduction in throughput, or an increase in latency (My assumption so far is the results are referring to throughput)

      I wasn't clear from the results whether this is the raw performance of the new socket classes assuming, or results obtained by actually transferring data on a real network.

      I guess my real question here is does this mean that if I use this transport then we can expect to be able to transport between 2 and 9 times less bytes / sec across a typical 100MB/s or gigabit network.

      JBossMessaing requires a UIL2 type transport, and this is what we're looking at using to provide this functionality, so I'm trying to work out whether I should be worried or not :)

      Thanks

      Tim

        • 1. Re: Multiplex transport performance.
          Ron Sigal Master

          Hi Tim,

          The performance numbers in the multiplex documentation are for a pair of virtual sockets running on localhost, and they represent the ratio of the time taken to transfer N bytes by virtual sockets compared to the time taken to transfer N bytes by real sockets. They're intented to give some idea of the penalty paid for multiplexing. Note that they're not based on the newest version of the multiplex system, and performance is improving.

          On the other hand, if you're interested in a multiplexed Remoting transport, then a comparison of the multiplex invoker to the socket invoker is more relevant, and the performance of the muliplex invoker is increasingly respectable. Earlier today I ran a test with 10 client threads making invocations on a server, and the multiplex invoker version took about 2.5 times as long as the socket invoker version. I expect the gap will continue to close.

          • 2. Re: Multiplex transport performance.
            Tim Fox Master

            Ok thx Ron-

            Let's see what performance figures we get in the perf lab then if it proves to be an issue we can address it then.