Performance analysis of wormhole switched interconnection networks with virtual channels and finite buffers

[摘要] An efficient interconnection network that provides high bandwidth and low latency interprocessor communication is critical to harness fully the computational power of large scale multicomputer. K-ary n-cube networks have been widely adopted in contemporary multicomputers due to their desirable properties. As such, the present study focuses on a performance analysis of K-ary n-cubes employing wormhole switching, virtual channels, and adaptive routing. The objective of this dissertation is twofold: to examine the performance of these networks, and to compare the performance merits of various topologies under different working conditions, by means of analytical modelling. Most existing analytical models reported in the literature have used a method originally proposed by Dally to capture the effects of virtual channels on network performance. This method is based on a Markov chain and it has been shown that its prediction accuracy degrades as traffic increases. Moreover, these studies have also constrained the buffer capacity to a single flit per channel, a simplifying assumption that has often been invoked to ease the derivation of the analytical models. Motivated by these observations, the first part of this research proposes a new method for modelling virtual channels, based on an M/G/1 queue. Owing to the generality of this method. Daily's method is shown to be a special case when the message service time is exponentially distributed. The second part of this research uses theoretical results of queuing systems to relax the single-flit buffer assumption. New analytical models are then proposed to capture the effects of deploying arbitrary size buffers on the performance of deterministic and adaptive routing algorithms. Simulation experiments reveal that results from the proposed analytical models are in close agreement with those obtained through simulation. Building on these new analytical models, the third part of this research compares the relative performance merits of K-ary n-cubes under different operating conditions, in the presence of finite size buffers and multiple virtual channels. Namely, the analysis first revisits the relative performance merits of the well-known 2D torus, 3D torus and hypercube under different implementation constraints. The analysis has then been extended to investigate the performance impact of arranging the total buffer space, allocated to a physical channel, into multiple virtual channels. Finally, the performance of adaptive routing has been compared to that of deterministic routing. While previous similar studies have only taken account of channel and router costs, the present analysis incorporates different intra-router delays, as well, and thus generates more realistic results. In fact, the results of this research differ notably from those reported in previous studies, illustrating the sensitivity of such studies to the level of detail, degree of accuracy and the realism of the assumptions adopted.

[发布机构] University:University of Glasgow

[关键词] Computer science