Throughput fairness in k-ary n-cube networks

Abstract

The performance of an interconnection network is measured by two metrics: average latency and peak network throughput. Network throughput is the total number of packets delivered per unit of time.Most synthetic network loads consist of sources injecting at the same given rate, using traffic patterns such as random, permutations or hot spot, which reflect the distribution of packet destinations in many parallel applications. The network is assumed to be fair: all source nodes are able to inject at the same rate. This work will show such assumption is unfounded for most router proposals. All router designs exhibited significant network unfairness under non-uniform loads. Some routers are also unfair under random traffic patterns. At loads above saturation, if the channel utilization is uneven, the injection matrix will become uneven: packet at low used areas will be accepted at a higher rate that those at the busy areas.As synthetic traffic does not reflect the coupled nature of the traffic generated by parallel applications, the impact of this unfairness on application performance could not be measured. New synthetic loads need to be developed to better evaluate network response beyond saturation.