Impact of Burst Size

In the previous experiment we did not analyze the effect of the burst size on the effective throughput. The burst size is the maximum number of new connections accepted concurrently for a given aggregate. With a large burst size, greedy clients can overload the server, whereas with a small burst, clients may be rejected unnecessarily. The burst size also controls the responsiveness of rate control. There is a tradeoff, however, between responsiveness and the achieved throughput.

We next show the effect of the burst size on the throughput of a preferred client. In our experiment, the non-preferred client is a modified sclient program that makes 50 to 80 back-to-back connection requests about twice a second, in addition to the specified request rate. Both the length of the incoming request burst and its timing are randomized. Figure 6 shows the throughput of preferred and non-preferred client with the SYN policing rate of the non-preferred client set to 50 conn/sec and the burst size varying from 5 to 50. The non-preferred sclient program requests a 16 KB dynamically generated cgi file. The preferred client is a Webstone program with 40 clients, requesting a static 8 KB file. As the burst size is increased from 5 to 50, the sclient's throughput increases from 36.6 conns/sec (585.6 KB/sec) to 47.7 conns/sec (752 KB/sec), while the throughput received by the preferred client decreases from about 140 conns/sec (1117 KB/sec) to 79 conns/sec.

Intuitively the overall throughput should have increased, however, the observed decrease in total throughput is due to the fact that we accept more CPU consuming CGI requests from sclient, thereby, incurring a higher overhead per byte transferred.

: