Utilizing spare network bandwidth to improve TCP performance
Manpreet Singh, Saikat Guha, Paul Francis · 2004
It is well-known that TCP Reno represents a performance bottleneck as the delay-bandwidth product increases. On high-bandwidth-delay links, the additive increase policy of one packet every RTT necessitates thousands of RTTs to reach full link utilization. This severely affects short TCP flows (most of the flows in the internet) as they cannot acquire bandwidth faster than “slow start ” and waste precious RTTs ramping up even when bandwidth is available. XCP [2] addresses this problem by heavily modifying routers to explicitly tell the end-hosts about their fair share of network bandwidth. Fast-TCP [1] backs-off proactively without incurring any loss but cannot quickly ramp-up as the fair share increases. We propose an approach where a TCP stream can quickly use the full available link bandwidth without requiring modifications to router software. Our approach rapidly fills the bandwidth-delay pipe while being just as cautious as TCP in avoiding causing losses to existing streams. This is accomplished by sending a large number of low-priority packets in addition to the conservative number of normal-priority TCP packets. Routers would be configured for strict priority queuing (already supported by routers) where low priority packets are routed only when no normal priority packets are in the queue, and normal priority packets are dropped only when there are no low priority packets in the queue. A diffserv-style packet marking scheme can be used to distinguish between normal-priority and low-priority packets.