In this section we present simulation results comparing the following mechanisms/policy combinations.
Figures 7 and 8 show the throughput results with the Apache and Flash Web servers, respectively, running on the back-end nodes. For comparison, results for the widely used Weighted Round-Robin (WRR) policy are also included, on HTTP/1.0 and HTTP/1.1 workloads.
When driving simple LARD with a HTTP/1.1 workload (simple-LARD-PHTTP), results show that the throughput suffers considerably (up to 39% with Apache and up to 54% with Flash), particularly at small to medium cluster sizes. The loss of locality more than offsets the reduced server overhead of persistent connections.
The key result, however, is that the extended LARD policy both with the multiple handoff mechanism and the back-end forwarding mechanism (multiHandoff-extLARD-PHTTP and BEforward-extLARD-PHTTP) are within 8% of the ideal mechanism and afford throughput gains of up to 20% when compared to simple-LARD. Moreover, the throughput achieved with each mechanism is within 6%, confirming that both mechanisms are competitive on today's Web workloads.
The performance of LARD with HTTP/1.1 (simple-LARD-PHTTP) catches up with that of the extended LARD schemes for larger clusters. The reason is as follows. With a sufficient number of back-end nodes, the aggregate cache size of the cluster becomes much larger than the working set, allowing each back-end to cache not only the targets assigned to it by the LARD policy, but also additional targets requested in HTTP/1.1 connections. Eventually, enough targets are cached in each back-end node to yield high cache hit rates not only for the first request in a HTTP/1.1 connection, but also for subsequent requests. As a result, the performance approaches (but cannot exceed) that of the extended LARD strategies for large cluster sizes.
WRR cannot obtain throughput advantages from the use of persistent connections on our workload, as it remains disk bound for all cluster sizes and is therefore unable to capitalize on the reduced CPU overhead of persistent connections. As previously reported [23], simple-LARD outperforms WRR by a large margin as the cluster size increases, because it can aggregate the node caches. With one server node, the performance with HTTP/1.1 is identical to HTTP/1.0, because the back-end servers are disk bound with all policies.
The results obtained with the Flash Web server, which are likely to predict future trends in Web server software performance, differ mainly in that the performance loss of simple-LARD-PHTTP is more significant than with Apache. This underscores the importance of an efficient mechanism for handling persistent connections in cluster servers with content-based request distribution.
The throughput gains afforded by the hypothetical ideal handoff mechanism might also be achievable by a powerful relaying front-end (see Section 3.1) as long as it is not a bottleneck. However, as shown in Figures 7 and 8, such a front-end achieves only 8% better throughput than the back-end forwarding mechanism used with the extended LARD policy.
