A simple and efficient mechanism to prevent saturation in wormhole networks

Abstract

Both deadlock avoidance and recovery techniques suffer from severe performance degradation when the network is close to or beyond saturation. This performance degradation appears because messages block in the network faster than they are drained by the escape paths in the deadlock avoidance strategies or the deadlock recovery mechanism.Many parallel applications produce bursty traffic that may saturate the network during some intervals [14,8], significantly increasing execution time. Therefore, the use of techniques that prevent network saturation are of crucial importance. Although several mechanisms have been proposed in the literature to reach this goal, some of them introduce some penalty when the network is not fully saturated, require complex hardware to be implemented or do not behave well under all network load conditions. In this paper, we propose a new mechanism to avoid network saturation that overcomes these drawbacks.

Exploiting global knowledge to achieve self-tuned congestion control for k-ary n-cube networks

“…For comparison, we also simulate the At-Least-One (ALO) [2] congestion control scheme which has been shown to be better than other previous schemes such as [22], [23]. ALO estimates global network congestion locally at each node.…”

“…Lopez et al [22], [23] use the number of busy output virtual channels in a node to estimate congestion. Baydal et al [2] propose an approach that counts a subset (free and useful) of virtual channel buffers to decide whether to throttle or not. Because the above schemes rely on local symptoms of congestion, they lack knowledge about the global network state, and are unable to take corrective action in a timely manner.…”

“…By limiting packet injection, our scheme sustains high throughput and low latency. Compared to an alternative approach that uses local estimates of congestion [2], our scheme is superior because global congestion estimation enables our technique to detect congestion in its early stages. We also show that a single static threshold cannot accommodate all communication patterns because a single threshold overthrottles some workloads and does not prevent saturation in other ones.…”

BLAM: a high-performance routing algorithm for virtual cut-through networks

“…This causes a sudden, severe drop in throughput and corresponding increase in packet latency. Several proposed schemes prevent the performance degradation that accompanies saturation by throttling packet injection when saturation is imminent [1,16,20,25,30]. Our approach goes further as it increases the applied load at which saturation occurs and avoids throughput degradation at saturation without placing limits on packet injection beyond those imposed by simple back-pressure.…”

“…Each packet is composed of flits (flow control units) that are transferred between network nodes. 1 We can classify routers as minimal or non-minimal. A minimal router offers only profitable routing options (hops that take a packet closer to its destination) whereas a nonminimal router may offer misrouting options (hops that take a packet farther from its destination) as well.…”