Microbenchmarks

All measurements below were taken on a set of Pentium II 450Mhz, 128 MB nodes, running Windows 2000 and connected using 100 Mbps switched Ethernet. Figure 6 shows the overheads introduced by CANS, measured in terms of how they impact communication between an application and an end service. Each graph shows the round-trip time and bandwidth achievable for different message sizes for four configurations: C prog and Java prog refer to our baselines, corresponding to application and server programs that communicate directly using native sockets in C or Java respectively. In process Driver and One EE refer to basic CANS configurations; the former shows the case when null drivers and a communication adaptor are embedded into the application interception layer and indicates the basic overheads of driver composition, and the latter considers the case where the data path includes null drivers on an intermediate host between the application and service. Figure 6 shows that the In process Driver configuration introduces minimal additional overheads when compared with the Java prog configuration (less than 10% arising from extra synchronization and data copying), attesting to the efficiency of our driver design and composition mechanism. On the other hand, the One EE configuration does show marked degradation in performance, primarily because of context switch costs and the fact that the transmitted data has to traverse across application-level and network-level four times instead of two times. However, given that intermediate EEs are intended to be used across different network domains where other factors dominate latency and bandwidth, this overhead is unlikely to have much overall impact.