In this section we evaluate the performance of the Slite NFS storage router. We compare an Anypoint Slite storage router (described in detail in Section 3.2) to user and kernel-level proxy alternatives.
The kernel-level (fastpath) configurations employ an Anypoint ALRM module to redirect frames. The ALRM sends most frames directly to servers; some require additional processing and go to the coordinator, which is just another server to the Anypoint layer. This category uses either an ACP or UDP transport. The UDP fastpath uses a simple kernel hook to intercept UDP packets and advertise them via the Anypoint interface to the Slite ALRM.
We use one coordinator proxy and four back-end NFS servers, exporting independent drives for a total of 32 NFS volumes (unified into a single logical volume). All machines run FreeBSD 4.4. Delayed acks are disabled to get good performance from the NFS RPC protocol over TCP, and the system is configured to use 8KB transport frames.
Figure 9 plots results from these four configurations, using the Fstress NFS benchmark [3] configured to generate a Web server's file system load. We measure average operation response time and CPU utilization at the switch with increasing request rate, plotted on the left and right, respectively. The TCP proxy has the lowest peak throughput and becomes CPU saturated at 5500 ops/s. The Anypoint switch shows a 29% improvement in throughput relative to the TCP proxy.
Note that the user-level UDP proxy gives better response times than the Anypoint switch, even though the Anypoint CPU overhead (shown in the right graph of Figure 9) is lower at any given request rate. We believe Anypoint/ACP is limited by the interaction between cumulative acks and congestion control described in the previous section.