OK. I can see the focus on cluster rollover. ;) We have conducted tests in a similar manner to queuing requests remotely. You can create a stateless EJB to make remote and local calls to itself (or other EJBs) as well as naming server lookups. If you look at our JBoss performance paper you'll see the code, which is quite simple http://www.amitysolutions.com.au/downloads/JBoss_code.pdf. So those kinds of things are nice and easy to do - since such a bean operation could be quite long running loop through and make 100,000 invocations, you could stack it up and get quite a few running to see the internal performance - if that is useful. You probably don't want so much load that CPU saturation deforms the response.
The problem for clustering will be to define a controlled test that is repeatable and easily configured for load whilst avoiding any deformations in the load generator. I think you will want to look at steady state responses and various degrees of load. How does round-robin perform with light, medium and heavy load? Versus other schemes? The bean method (internal load) will also need to be thought through. If it does a lot of processing, is this indicative of a real load - if an EJB spends time retrieving from a database, this results in lighter CPU load in the EJB container. If you introduce a database, how much influence will the connection pool operation effect times. Still thinking on this. I suspect as part of the tool set, you would need to build a JDBC loopback plug with a dummy return and tuneable response? This could be used to advantage to see how the app server responds to a pileup of waiting EJBs. I think some broad categories of tests need to be defined first and then refined. It will give a point of reference for devising these torture tests and how they would work. So let me summarise to date: Clustering Response - lookup? - purely invocation? - fixed computing load? - fixed time load? Aim: tabulate useful characteristics against single server versus multiserver with different load balancing schemes Issues: thread measurement, attaining steady state, load increment, deformation in responses due to measurement hooks, calibration of test rig and minimisation of non-linear effects in test rig (if they exist) or normalisation of results to account for this - depends on whether this is relative comparisons or standalone measurements. Regards, JonB
smime.p7s
Description: S/MIME cryptographic signature
