SpecWeb99 workload

On the SpecWeb99 workload, we find that mean response time is reduced by a factor of four by our changes. The cumulative distribution of latencies can be seen in Figure 10. We use 300 simultaneous connections, and compare the new server with the original Flash running on a patched VM system. Since 30% of the requests are for longer-running dynamic content, we also test the latencies of a SpecWeb99 test with only static requests. The mean of this workload is 7.1 msec, lower than the 10.6 msec mean for the new server running the complete workload. This difference suggests that further optimization of dynamic content handling may lead to even better performance. To compare the difference between static and dynamic request handling, we calculate the 5$^{th}$, 50$^{th}$, and 95$^{th}$ percentiles of the latencies for requests on the SpecWeb99 workload. These results are shown in Table 6, and indicate that dynamic content is served at roughly half the speed of its static counterpart. The latency difference between the new server and the original Flash on this test is not as large as expected because the working set still fits in physical memory.

Figure: Latency summary for 300 SpecWeb99 connections

Table: Separating SpecWeb99 static and dynamic latencies

	5%(ms)	50%(ms)	95%(ms)	mean(ms)
static	0.51	1.45	59.81	9.92
dynamic	0.99	2.83	91.31	12.19