1 Reply Latest reply: Aug 4, 2010 4:30 PM by Brian Oliver RSS

    Coherence with GPGPU

      Hi all,

      I am new to Coherence here.
      I read that most computation in Coherence were small task that needed to be processed quickly.

      Is there any real life scenario where we have some heavy computation in Coherence task so that we can submit to GPGPU accelerator for faster processing and then resend it back?

      ps: we can use CUDA with Java through JCUDA library - jcuda.org

        • 1. Re: Coherence with GPGPU
          Brian Oliver
          There are many possible ways you could achieve what you're thinking with Coherence. I all depends on the availability, scalability and performance requirements you require, as all of these things are trade-offs.

          Here are some suggestions, but with more information I'm sure we can narrow down the choices for your use case.

          1. Use Coherence EntryProcessors - to process individual or multiple entries in parallel.
          2. Use Coherence Invocation Service - to target processing at individual nodes in the cluster.
          3. Use the Incubator Command Pattern - to target processing at a context (and maintain order of execution for the said context)
          4. Use the Incubator Functor Pattern - much like the Command Pattern, but with support for asynchronous response.
          5. Use the Incubator Processing Pattern - very flexible approach to doing "processing" in a cluster. You could create you're own type of Dispatcher using JCUDA that would call out to CUDA.

          Hope this helps.

          -- Brian

Brian Oliver | Architect | Oracle Coherence