Check out the new USENIX Web site. next up previous
Next: Simulation Up: Efficient Support for P-HTTP Previous: The extended HTTP/1.1 LARD

   
Performance Analysis of Distribution Mechanisms

This section presents a simple analysis of the fundamental performance tradeoff in the use of the multiple handoff mechanism versus the back-end forwarding mechanism for request distribution in the presence of persistent connections.

When compared to the multiple handoff mechanism, the back-end forwarding mechanism trades off a per-byte response forwarding cost for a per-request handoff overhead. This would suggest that back-end request forwarding might be most appropriate for requests that result in small amounts of response data, while the multiple handoff approach should win in case of large responses, assuming that all other factors that affect performance are equal.


  
Figure 5: Apache
\begin{figure*}
\centerline{\psfig{figure=fig/apache_rvb.ps,height=3.0in}}\end{figure*}


  
Figure 6: Flash
\begin{figure*}
\centerline{\psfig{figure=fig/flash_rvb.ps,height=3.0in}}\end{figure*}

Figures 5 and 6 show the results of a simple analysis that confirms and quantifies this intuition. The analysis predicts the server bandwidth, as a function of average response size, that can be obtained from a cluster with four nodes, using either the multiple handoff or the back-end forwarding mechanism. The analysis is based on the values for handoff overhead, per-request overhead, and per-byte forwarding overhead reported above for the Apache and Flash Web servers, respectively.

To expose the full impact of the mechanisms, pessimal assumptions are made with respect to the request distribution policy. It is assumed that all requests after the first one arriving on a persistent connection have to be served by a back-end node other than the connection handling node. Since most practical policies can do better than this, the results indicate an upper bound on the impact of the choice of the request distribution mechanism on the actual cluster performance.

The results confirm that for small response sizes, the back-end forwarding mechanism yields higher performance, while the multiple handoff mechanism is superior for large responses. The crossover point depends on the relative cost of handoff versus data forwarding, and lies at 12 KB for Apache and 6 KB for Flash. These results are nearly independent of the average number of requests received on a persistent connection. Since the average response size in today's HTTP/1.0 Web traffic is less than 13 KB [19,3], these results indicate that the back-end forwarding mechanism is indeed competitive with the TCP multiple handoff mechanism on Web workloads.


next up previous
Next: Simulation Up: Efficient Support for P-HTTP Previous: The extended HTTP/1.1 LARD
Peter Druschel
1999-04-27