A methodology to enable QoS provision on InfiniBand hardware

Abstract

Modern high-speed interconnection networks include support for the provision of quality of service (QoS) to the applications. The output scheduling algorithm plays an important role in the QoS provision, choosing the packets to be delivered from the output buffers. InfiniBand, one of the most used interconnection technologies, includes a table-based scheduler composed of a high- and a low-priority tables, and a counter limiting the number of high priority traffic flows that may be delivered before giving the opportunity to low priority ones. Therefore, the performance of the traffic flows in the network largely depends on the table configuration since the switch scheduler uses this information to allow/deny packets being forwarded, according to the QoS provision scheme. As far as we know, there is no study on the influence of these configurations to the traffic flows performance. In this paper, we present an offline analysis tool to accurately determine the expected end-to-end latency and bandwidth of the traffic flows in an InfiniBand-based network using the information contained in the high- and low-priority tables. Moreover, we present a methodology to aid network administrators in configuring the QoS provision in a real InfiniBand cluster. Finally, we evaluate the analysis tool, comparing its results with those obtained from a real cluster and from simulation.