Figure 8: Web server throughput under bursty conditions versus request rate
In Section 3, we point out that one of the drawbacks of the naive traffic generation scheme is the lack of burstiness in the request traffic. A burst in request rate may temporarily overload the server beyond its capacity. Since Figure 7 indicates degraded performance under overload, we were motivated to investigate the performance of a Web server under bursty conditions.
We configured a S-Client with think times values such that it generates bursty request traffic. We characterize the bursty traffic by 2 parameters, a) the ratio between the maximum request rate and the average request rate, and b) the fraction of time for which the request rate exceeded the average rate. Whenever the request rate is above the mean, it is equal to the maximum. The period is 100 seconds. For four different combination of these parameters we varied the average request rate and measured the throughput of the server. Figure 8 plots the throughput of the Web server versus the average request rate. The first parameter in the label of each curve is the factor a) above, and the second is factor b) above, expressed as a percentage. For example, (6, 5) refers to the case where for 5% of the time the request rate is 6 times the average request rate.
As expected, even a small amount of burstiness can degrade the throughput of a Web server. For the case with 5% burst ratio and peak rate 6 times the average, the throughput for average request rates well below the server's capacity is degraded by 12-20%. In general, high burstiness both in parameter a) and in b) degrades the throughput substantially. This is to be expected given the reduced performance of a server beyond the saturation point in Figure 7.
Note that our workload only approximates what one would see on the real WWW. The point of this experiment is to show that the use of S-Clients enables the generation of request distributions of complex nature and with high peak rates. This is not possible using a simple scheme for request generation. Moreover, we have shown that the effect of such burstiness on server performance is significant.
