Web server benchmarking in a LAN environment fails to evaluate an important aspect of real Web workloads, namely that fact that clients contact the server through a wide-area network. The limited bandwidth and packet losses of a WAN increase the average HTTP connection duration, when compared to LAN environment. As a result, at a given throughput in requests/second, a real server handles a significantly larger number of concurrent connections than a server tested under LAN conditions [24].
The number of concurrent connections can have a significant impact on server performance [4]. Our next experiment measures the impact of the number of concurrent HTTP connections on our various servers. Persistent connections were used to simulate the effect of long-lasting WAN connections in a LAN-based testbed. We replay the ECE logs with a 90MB data set size to expose the performance effects of a limited file cache size. In Figure 12 we see the performance under Solaris as the number of number of simultaneous clients is increased.
 |
The SPED, AMPED and MT servers display an initial rise in performance as the number of concurrent connections increases. This increase is likely due to the added concurrency and various aggregation effects. For instance, a large number of connections increases the average number of completed I/O events reported in each select system call, amortizing the overhead of this operation over a larger number of I/O events.
As the number of concurrent connections exceeds 200, the performance of SPED and AMPED flattens while the MT server suffers a gradual decline in performance. This decline is related to the per-thread switching and space overhead of the MT architecture. The MP model suffers from additional per-process overhead, which results in a significant decline in performance as the number of concurrent connections increases.