If you are running a 100 Mbps network, full duplex and no collisions must be ensured for the network configuration. Note that some switches may require you to reboot for some configurations to take effect. In the event that a 1Gbps network is not available, a configuration can also be set up using cables to directly connect each pair of computers participating in the workload. As part of this study, we compared setups using direct cables and a 1Gbps switch, and we found no appreciable performance difference between the two scenarios.

While a 1Gbps network increases bandwidth and reduces latency, not all network drivers and configurations are optimal for the performance of a workload. For example, the Intel® PRO/1000 MT Server Adapter tends to reduce the number of interrupts but also supports configurations for more responsive interrupts, larger frame sizes, and so on. A performance engineer should examine the characteristics of these configurations to evaluate how applicable they are to the workload. In our study, we found that the default options for the network adaptor support reasonable performance.

The major performance statistics we examine for network-related performance and capacity are read/write (bytes/sec and packets/sec) and interrupts/sec. We used PERFMON to collect the values of these counters. As expected, a high ratio between bytes and packets tends to yield better network performance. However, the situation with network interrupts is not as clear, as a trade-off exists between response time and resource utilization. Another network statistic to check is that the total bandwidth requirement is within the capacity of the network link. One has to make sure that the network is not a bottleneck to ensure that performance analysis on the application server is meaningful.