Hi Folks,

I'm sorry these have been so long in coming. Attached is an HTML file of
the notes I took during the forum working group meetin at the beginning
of the month. They amount to basically a copy of what Dennis wrote on
the board. I believe George took a separate note of actions although
there are some peoples names down in this next to areas of
responsibility.

In terms of the benchmarking EPCC is going to provide some effort to
make a start on this before Supercomputing. We'll be contacting those of
you with you're names against benchmarks soon.

Thanks,

Martin



                           Applications and Concurrency

Chair: Dennis Gannon

Notes taken by: Martin Westhead

 

These notes were taken at the applications and concurrency working group of the Java Grande Forum on the 7th-8th of August
1998. A separate action list was taken by George K Thiruvathukal.

RMI performance issues

   1.RMI

Issues

          Support for non-TCP & HP protocols (Memory, ATM, Myranet, FM…,IPC) [API] 
          Connection management (too frequently opening & closing connections)[IMP] 
          Class compatibility [IMP] 
          Resource consumption (implementation issue)[IMP] 
          Scheduling conflicts [IMP]

 

Solutions

          Possible soln GMI a new object 

 

   1.Serialization

 

Issues

     Too much type information 
     Efficient support for non-recursive (flat) structures [API] 
     Buffer size control [IMP] 
     Too many copies [IMP] 
     Serialization of float & double goes through the JNI 
     Serialization of objects with multiple primitive instance variables is inefficient because it requires individual expensive calls
     for each variable

Solutions

     Serialization API 
     Caching and connection persistence 
     Aggregate serialization of objects with multiple primitive instance variables 

VM issues

     Process/Threads: Affinity/co-scheduling 
     Memory management : 
          How does the VM optimize deep memory hierarchies? 
          Large objects? Limitations 
          Large number of objects
     Staging large objects (3d graphics problem) 
     Scalable I/O

Benchmarks (Martin Westhead)

Aim to find programs that will stress the VM in ways that are important to Grande application

(Check out netlib benchmarks. Talk to Java 3D)

 

Applications

          Monte Carlo (NCSA + EPCC: financial) 
          Image analysis/Radio astronomy (NCSA) 
          Gravitational N-body Simulation 
          Computational Fluid Dynamics (Syracuse) 
          Geophysics (Michael) 
          Discrete Event Simulation (INRIA + EPCC)

Kernels

          Spec Java Kernels 
          Mpeg3 audio decoder 
          Javac 
          Jess 
          Ray tracing 
          Sorting – George 
          Threads – synchronization, scalability, threads v. runable – Dennis 
          RMI – Michael 
          Serialization - Michael 
          I/O – Dan Reed 
          Memory management – object size, object number, garbage collection - PM 
          Sun benchmarks? 
          Low level arithmetics – Numerics Group 
          Linpack - NG 
          Jama – NG

Comparisons (Email from Vladimir)

          JVMs 
          Compilers v. JVMs 
          Language comparisons

Seamless comp (PM)

Define a simple common interface: java.lang.org.grande.seamless

          Relationship to other technologies: 
          Unicore 
          Globus 
          Legion 
          Condor 
          Sweb 
          Websubmit 
          Issues 
          Authentication 
          Resource discovery (metadata) 
          Resource allocation (Grande cache) 
          Accounting 
          Resource and job Scheduling 
          Job monitoring 
          Code/Data-filesystem 
          Connected to resource discovery service (JINI) 
          Action will involve an examination of the various technologies in order to refine issue list

 

Other parallel issues 

          MPI 
          Pure bindings 
               Interface style issue 
               Not feasible or sensible to do whole MPI 
               Request for Proposals for a deprecated MPI interface 
               Issues with heterogeneous language interaction
          Extended MPI flavour 
               JMPI 
               Difficult to reach a consensus 

                    Could be built on top of a pure MPI 
                    Too early to produce a specification but encourage experimentation in this area

          Object model : value semantic in proc calls, remote reference management, problems for large objects [API] –
          Controversial and moved from RMI section 
          Threads issues