Introduction
This report presents performance data for a test that compares the effects of WebSphere® Application Server vertical JVM stacking to horizontal JVM stacking.
Customers have often asked for guidance on how best to implement a WebSphere Application Server multiple server environment. The question is whether it is better to host all WebSphere Application Servers on a single (vertical JVM1 stacking) z/VM® guest, or distribute the servers across multiple guests (horizontal JVM stacking). In the past, it has been found that two portal servers on the same z/VM guest provided better performance than having them distributed across two z/VM guests.
- The management overhead in z/VM scales with the number of guests.
- The management overhead and resource competition in Linux™ scales with the number of JVMs per Linux instance.
This means, for example, that having all JVMs in one Linux guest provides the lowest z/VM overhead, but the highest overhead in Linux. But one JVM per guest and many guests result in the opposite behavior. The question is which of these mechanisms has the higher impact, and are there are other, currently unknown, factors that also have an impact on the system. The expectation is that the optimal scenario is somewhere in between the extreme ends.
- All JVMs in one guest (full vertical stacking)
- One JVM in each guest, and many guests (full horizontal stacking)
- Configurations in between
The workload used in driving the various WebSphere Application Server environments was a simple web services application that did not require any database.
The number of JVMs (200) used in the test was held constant. With only one guest in the configuration, the 200 WebSphere Application Servers (JVMs) are defined on one guest. With two guests, each guest runs 100 WebSphere Application Servers. As the number of guests are scaled, the number of WebSphere Application Servers (JVMs) defined per guest decreases. However, the total number of WebSphere Application Servers (JVMs) under test is always held at 200.
As an additional test, the use of shared mini-disks or a DCSS for the WebSphere installation was analyzed, to determine if there are criteria that favor one or the other setup.