next up previous
Next: Conclusion Up: Experimental Results Previous: TCP Overhead

UDP Latency

Figure 6: UDP One-Way Message Latency
\epsfig {file = figs/latency.epsi, scale = .45, angle=-90, width=\linewidth}

Figure 6 shows the one-way UDP latency for various packet sizes using Trapeze and the Alteon Gigabit Ethernet on Monet. These results were obtained with netperf -tUDP_RR -l60, which ping-pongs a packet of the requested size for one minute. We plotted points for one half of the average round-trip latency for one-byte packets and for packet sizes of 1KB to 8KB in 1KB increments. For these experiments we disabled checksum offloading on the Alteon after observing unexpectedly high latencies with checksum offloading enabled. The comparison is fair because Trapeze does not support checksum offloading on this platform. The two sets of lines for each configuration show the latency with software checksums and with checksums disabled. Zero-copy sockets are responsible for the lower latency in all configurations when the 8KB page size is reached.

In the Trapeze runs, the slope change at 1KB results from use of a payload buffer. The one-byte Trapeze packets are sent in a control message with no payload buffer, resulting in lower latency. For the 1KB-7KB sizes, the data is copied into a payload buffer and sent as a Trapeze payload.

The latency results show the benefits of message pipelining in Trapeze, which overlaps transfers on the link with transfers on the sender and receiver I/O bus. This overlap causes packet latencies to grow at a slower rate as packet size increases. In fact, the experiment understates these benefits because message pipelining is supported only on the receiver on the LANai-5 NIC used in this configuration, due to a change in the meaning of certain control registers on the LANai-5.

next up previous
Next: Conclusion Up: Experimental Results Previous: TCP Overhead
Jeff Chase