Tuning SPECjAppServer2002

Components

Hardware

Single or multiple hardware servers can be involved in a SPECjAppServer2002 benchmark. The initial sizing of the servers is very important as well as having the capacity to grow. Understanding the growth requirements for any application can be difficult if no data exists on how the application behaves under different workloads and stress levels. The following suggestions should be kept in mind, when running SPECjAppServer2002:

1.      Either Task Manager or PERFMON can be used to monitor the CPU usage of a benchmark run. If the average CPU utilization over the steady state run of the SPECjAppServer application of the monitored system exceeds 80%, it&146;s probably time to increase the number of CPUs. This type of monitoring should be conducted on all servers participating in the benchmark and should be cyclic. For example, in one benchmark run it might be determined that more CPUs should be added to the server running the SPECjAppServer2002 Application. However, increased capacity on that server may cause increased demands on the server running the Database. The next benchmark run may identify this new bottleneck, allowing more CPUs to be added to the server running the Database.

2.      When using multiple servers, the highest specification Network Interface Cards (NIC) should be used e.g. 1 Gigabit. If 1 Gigabit NICs are not available, 100Mbit NICs can be used instead, but the throughput may not be as good. Each server should also be connected to each other via a private network. A network accessible by other users e.g. a public network could be used, but this will affect throughput. In the unlikely event that server manages to exceed the capacity of a 1 Gigibit NIC (can be monitored with PERFMON), additional NICs could be added into the server. For example, if the server running SPECjAppServer2002 Application, which has to talk to the Benchmark Driver, SPECjAppServer2002 Emulator and Database, could have 3 NICs installed &150; one to deal with each other external server.

3.      Having good NICs is only part of the network story. Attention also has to be paid to network routers and their configuration.

4.      The server running the Database should be configured with fast disk drives (10,000 RPM and above), preferably in a RAID array. PERFMON can be used to monitor the disk activity. Excessive Average Read/Write Queue lengths could mean that either additional drives or drive controllers need to be added or more memory needs to be given to the server running the Database. Servers running other SPECjAppServer2002 components are not as sensitive to disk activity as the server running the Database. Typically, the average disk write and read response time should be less than 10ms.

5.      Each server should be installed with as much memory as possible.

Referenced by: