We used a segment of the Rice University trace alluded to in Section 6 to drive our prototype cluster. A single back-end node running Apache 1.3.3 can deliver about 151 req/s on this trace.
The Apache Web server relies on the file caching services of the underlying operating system. FreeBSD uses a unified buffer cache, where cached files are competing with user processes for physical memory pages. All page replacement is controlled by FreeBSD's pageout daemon, which implements a variant of the clock algorithm [18]. The cache size is variable, and depends on main memory pressure from user applications. In our 128MB back-ends, memory demands from kernel and Apache server processes leave about 100MB of free memory. In practice, we observed file cache sizes between 70 and 95 MB.
The mechanism used for the WRR policy is similar to the simple TCP handoff in that the data from the back-end servers is sent directly to the clients. However, the assignment of connections to back-end nodes is purely load-based.
Several observations can be made from the results presented in Figure 13. The measurements largely confirm the simulation results presented in Section 6.1. Contrary to the simulation results, WRR realizes modest performance improvements on HTTP/1.1 on this disk-bound workload. We believe that HTTP/1.1 reduces the memory demands of the Apache server application, and therefore leaves more room for the file system cache, causing better hit rates. This effect is not modeled by our simulator.
The extended LARD policy with the back-end forwarding mechanism affords four times as much throughput as WRR both with or without persistent connections and up to 26% better throughput with persistent connections than without. Without a mechanism for distributing HTTP/1.1 requests among back-end nodes, the LARD policies perform up to 35% worse in the presence of persistent connections.
Running extended LARD with the back-end forwarding mechanism and with six back-end nodes results in a CPU utilization of about 60% at the front-end. This indicates that the front-end can support 10 back-ends of equal CPU speed. Scalability to larger cluster sizes can be achieved by employing an SMP based front-end machine.
