In this section, we present performance results obtained with our prototype cluster. Section 6 presents performance results of the software layer-4 switch discussed in Section 5.2. In Section 6.1, we present performance characteristics of our cluster prototype such as scalability, throughput and latency. Finally, Section 6.2 provides experimental results on a real workload obtained from webserver logs. For all cluster experiments we used the configuration shown in Figure 7. The experimental setup used was the same as described in Section 3.
The first experiment determines the maximum throughput of our software layer-4 switch used as a front-end. The switch receives packets from seven client machines. The packet processing in the switch closely emulates the processing in actual operation with the cluster. Each packet was forwarded to another machine where it was discarded.
Our results show that the peak throughput afforded by the switch running on a 300MHz Pentium II machine is about 128,000 packets/s. In actual cluster operation, we have observed an average of about 5-6 packets for a typical HTTP/1.0 connection where the content size ranges from 5-13 KB. This implies that the maximum throughput afforded by the switch is about 20,000 conn/s.
Higher switch performance can be obtained by (1) using faster hardware for the switch, (2) using an SMP based machine for the switch, or (3) by using a commercial layer-4 switch. Hardware implementations of layer-4 switches are reported to achieve throughputs of over 70,000 conn/s [1].