The ONL Tutorial

Tutorial >> More Plugins

TOC

Plugin Overhead

The SPC is a 500 MHz Pentium III mobile processor running a NetBSD kernel. In order to avoid overruning the SPC, the link rate to the SPC is limited to 200 Mbps in a standard ONL configuration. This page describes some measurements that indicate how we chose this limit.

In these experiments, the 200 Mbps limit was removed, and we sent iperf UDP packets to the SPC at an increasing rate until we encountered a bottleneck. We tried both large packets (1470 bytes or 31 ATM cells) and small packets (52 bytes or 2 ATM cells).

Fig. 1. SPC Throughput (Large Packets).

Fig. 1 shows the results of an experiment with large packets. The top graph shows that the ATM bandwidth peaks at around 200 Mbps. As the input rate is increased above 200 Mbps (black curve), the output rate from the SPC (blue curve) remains at 200 Mbps. The bottom graph shows that the peak packet processing rate of the SPC for large packets is around 14,000 packets per second.

Fig. 2. SPC Throughput (Small Packets).

Fig. 2 shows the results of an experiment with small packets. The top graph shows that the ATM bandwidth peaks at around 115 Mbps. As the input rate is increased above 115 Mbps (black curve), the output rate from the SPC (blue curve) remains at 115 Mbps. The bottom graph shows that the peak packet processing rate of the SPC for large packets is around 140,000 packets per second; i.e., 10 times the large packet experiment. But because the packets are more than 10 times smaller than the large packets (about 30 times smaller), the ATM throughput is less than 200 Mbps.

There are two types of overhead:

• A per packet overhead due to interrupt processing and passing control to the handle_packet routine of a plugin.
• A packet length overhead associated with the transfer of a packet to and from memory.

Tutorial >> More Plugins	TOC