A family of mechanisms for congestion control in wormhole networks

Abstract

Multiprocessor interconnection networks may reach congestion with high traffic loads, which prevents reaching the wished performance. Unfortunately, many of the mechanisms proposed in the literature for congestion control either suffer from a lack of robustness, being unable to work properly with different traffic patterns or message lengths, or detect congestion relying on global information that wastes some network bandwidth. This paper presents a family of mechanisms to avoid network congestion in wormhole networks. All of them need only local information, applying message throttling when it is required. The proposed mechanisms use different strategies to detect network congestion and also apply different corrective actions. The mechanisms are evaluated and compared for several network loads and topologies, noticeably improving network performance with high loads but without penalizing network behavior for low and medium traffic rates, where no congestion control is required.

Enhancing Entropy Throttling: New Classes of Injection Control in Interconnection Networks

“…The major approach is to throttle new packet injection. Many research results are reported in the literature [15]- [19]. All of the proposed throttling methods introduce additional information to start and to stop throttling.…”

Realistic Evaluation of Interconnection Network Performance at High Loads

“…As most network proposals sustained their maximum throughput after saturation, the evaluation of the network at heavy loads was not considered of interest until recently. However, congestion is a problem for large INs with multiple injection sources, and congestion control mechanisms are tested at loads beyond saturation [1,9,14] in order to see if they prevent traffic bursts from degrading network performance.…”

“…Besides, an unbalanced load distribution results in the formation of persistent zones of congestion at high loads. In this context, average throughput as reported in [1,8,15] is not representative of application loads, because "fast" injecting nodes will eventually wait for the slow ones to catch up. Therefore, in order to obtain useful performance figures at heavy loads, a static synthetic traffic pattern should reproduce the level of coupling that exists amongst traffic sources in actual parallel applications.…”

“…Non-static loads reflect the fact that communication in many parallel applications is not constant over time, so that an intensive communication phase (with high traffic volume) will be followed by a computation intensive phase (with low traffic volume), Recent studies using non-static traffic patterns include [1,14] in which traffic is uniform but load alternates between low and high phases, and [15] in which each high phase uses a different pattern. As mentioned before, the figure of merit is total execution time.…”

“…This is not the case for large networks with multiple injection sources such as IBM's BlueGene/L [2], which are prone to suffer from congestion at heavy loads. Consequently, new congestion control techniques have been proposed and evaluated for wormhole [1,15] and virtual cut-through INs [9,11,14]. Most of these works carry out evaluations using synthetic traffic patterns that assume that nodes generate traffic independently of each other.…”