Do the machines really have to be that large? Even WebSphere rarely actually requires 768M.
Overcommitting is good, it's the only way to make people share the resource. If you have enough resources to avoid sharing you don't need to encourage it. For WebSphere it obviously that depends on the application running there. And it depends on what else you run. Most of the measurements look at throughput in a single virtual machine, but when you have multiple guests running websphere your target may be different. Even when you find that it likes 768M (what a wonderful magic number that is) this may not be what is best to achieve maximum consistent throughput.
One of the ideas is that you should not mix applications in a Linux virtual machine because that makes it hard for VM to tell what the guest is doing and provide proper resources. And when you run WebSphere you really have 3 different applications in one virtual machine. The database will attempt to grab all memory to cache the database I/O. The Java processes tend to take a lot of CPU resources, but only once per transaction. The front-end web server gets involved several times for a single transaction (to deliver yet another chunk of http traffic, driven by the latency of the network). When you combine the Java processes with the front-end web server the virtual machine ends up in Q2 or Q3 because it uses a lot of CPU, but this also means it is dispatched less often. That's too bad because the web server only uses small amounts of CPU cycles and would benefit from being serviced quicker. A similar discussion applies to memory usage. When the database uses memory to cache data from the database, it takes away the free memory that the web server would otherwise have around to build buffers for the http data. The individual virtual machines can now be sized and tuned to meet the requirements.
In the "Performance and Tuning" Redbook SG24-6926 we did some work in this area and found we could at least get the same throughput in the configuration with 3 virtual machines (but maximum throughput is not the only issue). The recent patches to share userspace code in DCSS should help to cut down the cost of the additional base Linux systems, and make it even more attractive to break up the application.
Rob
---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
