Topology can have a significant effect on WebSphere performance. This article describes some of the topology considerations you should be aware of when configuring and installing WebSphere for z/OS.
WebSphere for z/OS gives you the ability to install your application either in a single server or spread it across multiple servers. There are many reasons for partitioning your application. However, for performance, placing your application all in the same server will always provide better performance than partitioning it. If you do choose to partition your application across servers, you will get better performance if there are at least replica servers on each system in the sysplex. The WebSphere for z/OS runtime will try to keep calls local to the system if it can, which will, for example, use local interprocess calls rather than sockets.
You also have a choice of running server regions with an isolation policy of one tran per server region or multiple trans per server region. From a performance perspective, running more threads in a server region will consume less memory but at the cost of thread contention. This contention is application-dependent. We generally recommend the use of multiple trans unless you run into contention problems.
Specify the threads setting using the server_region_workload_profile. The variables include:
Note: Please see the "Servers" section in the WebSphere Application Server for z/OS V5.0 information center, access to which can be obtained through the WebSphere forz/OS library Web site http://www.ibm.com/software/webservers/appserv/zos_os390/library.html for more information on ORB services advanced settings.
On a local client, the client and the optimized communication are done on the same system. This has some additional client CPU costs but less communication cost. On a remote client, the client cost is replaced by the additional communication overhead of sockets. The CPU cost on either system is almost equivalent. Latency is better for a local client than for a remote client, meaning you will get better response time with a local client.
You can define more than one copy of a server on a system. These copies are called clones. We have found slight improvements in performance when running with a couple of clones as opposed to just one (very large configuration). While there is some benefit, IBM does not recommend, at this time, the creation of replicated control regions for the sole purpose of improving performance. We do, however, recommend them for eliminating a single point of failure and for handling rolling upgrades without introducing an outage.