Scalability of Cougaar

[Todd Wright]

Dimitrios Traskas wrote:
> Hi
> 
> I am a new Cougaar developer and decided to move to the platform after 
> failed attempts with JADE and Agentscape when tried to scale my 
> projects. I am running large scale simulations that I wish to extend to 
> 2000+ Agents at some point. From your experience do you believe that 
> Cougaar is stable and scalable enough to handle thousands of Agents at 
> the same host and several hosts. How many Agents would you say it can 
> have in one machine? I would really appreciate your comments.

We've tested 5 JVMs with 2000 agents each, where the agents are configured into 
pairs that periodically "ping" one another.  The test worked but the naming 
service overhead was high.

You can create similar scalability test with the ping tutorial:
   http://tutorials.cougaar.org/
For example, you could easily write a script to generate an XML configuration 
for a fan-shaped ping (as seen in the "FanSociety.xml") with thousands of 
agents split across multiple nodes (as seen in the "PingTwoNodeSociety.xml"). 
Also see the "mesh" tutorial.

A single-host configuration can be optimized to use the local-only naming and 
messaging services, as illustrated in the above demos.

The multi-host naming service can be tuned or, if required, replaced with an 
alternative implementation.  For example, the pluggable "local-only" naming 
service implementation is about 200 lines of Java code:
http://cougaar.org/cgi-bin/viewcvs.cgi/core/src/org/cougaar/core/wp/LoopbackWhitePages.java?rev=1.2&cvsroot=core&content-type=text/vnd.viewcvs-markup

Todd

> 
> 
> Thanks
> Dimitri