ObjectWeb Consortium
Search ObjectWeb Mail Archive: 

Advanced Search - Powered by Google

Mail Archive Home | proactive List | May 2005 Index

<--  Date Index  --> <--  Thread Index  -->

Re: [proactive] non-linear performance

Hi Guillem,
guillem lefait wrote:

Hi there !
I'm working for my trainee on a TSP' solver. I got two problem/questions :
I got some pending connexion from worker to manager and theses one are
not killed at the end of application therefore i used an exit fonction
public void exit(){
   try {
        System.out.println("Kill les JVMs");
   } catch (Exception e) {
     System.out.println("Killing the application!");
So every time i launch an application on cluster, each node get an
additional rsh lauching process active.
=> What can i do for this ?
What is the job scheduler on the cluster. I don't know where this rsh process is from. Does the exit function work e.g, is the jvm killed ? If you are deploying on remote machines with rsh(no job scheduler), you can still "close" the rsh stream, so you won't have any rsh process left, but also you won't get any output from remote machines unless you configure the remote log4j file accordingly. To close the stream, just specify in the definition of your rsh process: closeStream="yes": For instance
<processDefinition id="rsh_amstel">
<rshProcess class="org.objectweb.proactive.core.process.rsh.RSHProcess" hostname="amstel.inria.fr" closeStream="yes">
                   <processReference refid="localJVM"/>
then the rsh streams (in/out/err) are closed as soon as the rsh command is performed

I've build a second version of my tsp'solver. i got this execution
time for a generic bench :
1 node : 2min 51
2nodes: 1min 34
3 nodes :1min 10
4 nodes : 55 sec.

My first version do the same in 50 sec on 2 nodes :(.
Can you confirm me that usually object are passed by deep copy and not
with stub ?
All objects are passed by deep-copy in the same jvm, by Serialization on different ones, except Active object that are passed as reference.

I don't really know why i got this powerfull lack - even if i know i'm
really not a java'guru :)-.
If you run your application on a single jvm, I would say it's normal, because of ProActive's overhead for the communications compare to normal java ones. On distributed systems, you will probably get speedup on larger systems. Anyway, we are working on ProActive's performances issues, to try minimizing the difference, with normal RMI communications.

If anyone want to check my appli, to help me increase perf or just for
looking at it, just send me a mail

Anyway, i will test my solver with a PVM way. I can send there result
even if Java can't really be compared to C.


Romain Quilici
Software Engineer
OASIS Team - INRIA Sophia-Antipolis
04 92 38 71 65

<--  Date Index  --> <--  Thread Index  -->

Reply via email to:

Powered by MHonArc.

Copyright © 1999-2005, ObjectWeb Consortium | contact | webmaster.