We next study the performance of TCP and PCP flows when they are transmitted through routers that implement fair queueing. Our goal is to show that PCP is compatible with and benefits from general-purpose network enforcement. Others have argued for TCP-specific rules for penalizing misbehaving endpoints [20], but we argue that this unnecessarily constrains the choice of end host algorithm.
|
In theory, fair queueing is neutral to the choice of endpoint congestion control algorithm. An endpoint that sends faster than its fair share will build up queues and cause packet loss, but only to its own packets. However, the inefficiency imposed by TCP slow start has no benefit if the bottleneck resource is fair queued, but the endpoint has no way in general of knowing how the network is being managed.
We use the same simulation parameters as those used in the previous experiment, but substitute the FIFO drop-tail router with a fair-queued router. Figure 5 shows that PCP benefits from fair queueing, with lower response time variance even at high loads. Bandwidth probes initiated by arriving PCP flows are allowed to succeed due to the isolation of flows by the fair-queued router. TCP, on the other hand, performs slightly worse with fair queueing. When a TCP flow exceeds its fair share of router buffer space under loaded conditions, it suffers multiple losses, causing its performance to drop severely; in contrast, a drop-tail router spreads the packet losses more smoothly across all of its flows.
