Sunday, July 06, 2003

I am still working on a n-machine cache system. Each machine holds a subset of the domain (a.k.a., domain prime) and can answer queries requiring a complex object graph (we expect 2 million queries a day soon). I am using a JBoss cluster for the query machines. The problem is that the client does not know which query boxes are running, so I configure the InitialContext of the client using cluster identity rather than machine identity. Therefore, any server in the cluster that is currently running can answer the query. This brings me to my point: JBoss clustering is based on javagroups. This is an extremely cool and apparently rock-solid api that implements group functionality based on multicasting and a configurable protocol stack. One of the included sample apps extends a HashMap with a DistributableHashMap. Think transparently-distributed local memory. When an element is put into the HashMap of one VM, it shows up magically on all the other VM's HashMaps that belong to the group. JavaGroups knows when new members join or leave the group. To better understand JBoss clustering, take a look at JavaGroups. JBoss points out that JavaGroups could be swapped out for another implmentation, but for me, studying JavaGroups made the JBoss clustering stuff click. 9:19:26 AM

© Copyright 2003 Larry Sherrill. Last update: 9/21/2003; 12:20:33 PM.