Routing and performance evaluation in interconnection networks

G. Stamoulis

Abstract:We analyze routing problems for message-passing parallel computers. First, we study static problems. We derive an algorithm for performing the total exchange task in the shortest possible time. We also derive an efficient algorithm for the task where each member of a subset of nodes wishes simultaneously to broadcast a packet to all other nodes. We then consider dynamic routing problems. The analysis constitutes the main focus of our research. Our motivation is the understanding of the communication issues arising in general purpose asynchronous computation. We analyze the problem of multiple node-to-node communications in the hypercube: each node generates packets according to a Poisson process; each packet has a single destination, which is selected randomly. We consider a simple greedy routing scheme, where every packet takes a particular shortest path leading from its origin to its destination. We analyze this scheme by treating the hypercube as a queueing network with deterministic servers, and by using a stochastic comparison with a product-form queueing network. We thus prove that the greedy scheme can sustain any throughput value less than 1, while inducing an average delay of $\Theta(d)$, (of the order of magnitude of the diameter $d$ of the hypercube) for any fixed throughput value. We also prove that, in heavy traffic, the average delay has optimal asymptotic behavior. We then extend these results to the butterfly network. Finally, we consider the dynamic routing problem of multiple broadcasts in the hypercube, where packets to be broadcast are generated according to Poisson processes. We devise and analyze two routing schemes that can sustain high throughput values regardless of the size of the hypercube; at the same time, in light traffic, the average time required to complete a broadcast is $\Theta(d)$, which is the optimal order of magnitude. The first routing scheme is analyzed exactly and closed form results are obtained. The performance of the second scheme is investigated using an approximate model; a simulation study shows that the resulting formulas are very accurate. All of the routing schemes considered in both dynamic problems are distributed and of the on-line type. (Abstract shortened with permission of school.) (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)