Overheads of the Web switch

To stress at most the Web switch, it is important that the Web servers of the cluster provide the fastest response possible. For this reason, the overhead analysis is carried out in a special scenario where the clients issue requests for one small-sized, static file that is always served by the disk caches of the server. We first measure the response time provided by a server that is directly connected to the client without Web switch. We then perform the same experiments by interposing the proposed ClubWeb-1w switch between the client and the server. Figures 5 and 6 show the mean response time and the throughput, respectively, as a function of the requested file size. We can conclude that the ClubWeb-1w overhead does not modify the throughput of the cluster, while it has a low impact on the response time as long as the network does not become the bottleneck of the whole system (that is, when the file dimension is lower than 5Kb). On a dual-CPU Web switch the maximum response time increases linearly up to 6.71% (at 3Kb). It is worth to note that with small file sizes that can fit into one IP packet, ClubWeb-1w does not show any overhead with respect to the single server case. Other tests for file sizes greater than 5KB saturated the network capacity of the FastEthernet, that was utilized at a peak rate of 89.1 Mbps.

Figure 5: Response time of ClubWeb-1w vs. single server.

Figure 6: Throughput of ClubWeb-1w vs. single server.