On Tuesday, October 28, 2003, at 02:02 AM, Jules Gosnell wrote:
Bhagwat, Hrishikesh wrote:
the impact of moving large number of session objects on the network is a very obvious concern (expressed in the mail thread below)if network bandwidth is the bottleneck, then compression is a possible solution. Bear in mind that you are trading cpu and memory requirements for bandwidth.
demarshalling is an expensive thing to do and will affect backup nodes as well as target nodes in migration relationships.
if a backup node can avoid demarshalling incoming objects, then this is a worthwhile optimisation.
compression, which is used in JBoss(tm) SFSB replication is a useful weapon in the arsenal.
I read several research papers on this subject and all seem agree that network bandwidth is never the problem (unless your cluster is over a wan). The bottle neck tends to be the IO processing. Basically, you run out of CPU trying to respond to the other nodes lock requests or data replication requests. Anything you can do to reduce CPU overhead will improve scalability on the cluster. Compression and encryption are in my opinion a very bad for the back end of the cluster as bandwidth inside your data center should be very cheap. On the front end compression and encryption are commonly requirements, but can be off loaded to a speciality box (load balancer).
-dain
