Impact Of Network Fairness On The Performance Of Parallel Systems

Execution time of parallel applications depends on the balanced execution and synchronization of all its processes. However, most interconnects exhibit significant throughput unfairness, introducing load unbalance. At high loads, such unfairness significantly degrades the performance of some nodes, and eventually the whole system. Different strategies have been developed to guarantee network fairness, fundamentally based on two different approaches: traffic prioritization in the network switches or injection throttling at the source. This work qualitatively and quantitatively compares two fairness mechanisms which are representative of these two types of explicit fairness strategies: Age arbitration, which prioritizes old packets in transit, and the SAT injection protocol, which throttles traffic to provide even access to all nodes. Multiple synthetic traffic patterns as well as parallel application loads at different transmission rates are considered in the evaluation.Our research identifies key limitations of each fairness mechanism: a realistic implementation of Age arbitration fails to achieve perfect fairness due to a delayed assignment of packet timestamps, while SAT may restrict maximum throughput when it is not necessary to do so. Both mechanisms do not hinder performance at low loads, and improve throughtput fairness at congested loads. With real applications, performance improvements up to 14.2% are observed, with SAT being particularly efficient at high loads.