Adaptive Multipath Routing for Congestion Control in InfiniBand Networks

摘要

InfiniBand networks are aimed to fulfill increasing communication demand of parallel applications in High Performance Clusters. However, the dynamic behavior of communication load may cause bottlenecked link(s) which lead to message congestion. Congestion spreading increases latency and reduces network throughput causing important performance degradation. In this paper, we propose a congestion control technique for InfiniBand networks. Our proposal is based on an adaptive routing algorithm that performs a communication load balancing over several alternative paths. Congestion is controlled by avoiding sending communication-load excess through network congested area, while maintaining injection rate. Proposed mechanism uses the management model defined in InfiniBand specs, thus full compatibility is provided meanwhile congestion control is addressed.