Improving the Up*/Down* Routing Scheme for Networks of Workstations

Abstract

Networks of workstations (NOWs) are being considered as a costeffective alternative to parallel computers. Many NOWs are arranged as a switchbased network with irregular topology, which makes routing and deadlock avoidance quite complicated. Current proposals use the up * /down * routing algorithm to remove cyclic dependencies between channels and avoid deadlock. Recently, a simple and effective methodology to compute up * /down * routing tables has been proposed by us. The resulting up * /down * routing scheme makes use of a different link direction assignment to compute routing tables. Assignment of link direction is based on generating an underlying acyclic connected graph from the network graph. In this paper, we propose and evaluate new heuristic rules to compute the underlying graph. Moreover, we propose a traffic balancing algorithm to obtain more efficient up * /down * routing tables when source routing is used. Evaluation results show that the routing algorithm based on the new methodology increases throughput by a factor of up to 2.8 in large networks, also reducing latency significantly.

Deadlock-free routing based on ordered links

“…The effect of sorting the links (d BF OT + − d BF OT ) seems to be more significant than the difference between BFOT and APOT. Similar effects are observed when picking a different node as the root in the up*/down* algorithm, as noted in [9].…”

Optimising Up*/Down* Routing By Minimal Paths

“…The selection of the root is made by using heuristic rules [8]. For instance, the switch with the highest average topological distance to the rest of the switches will be selected as the root node.…”

“…The rest of the switches are added to the DFS spanning tree following a recursive procedure. Unlike the BFS spanning tree, adding switches is made by using heuristic rules [8]. Starting from the root switch, the switch with the highest number of links connecting to switches that already belong to the tree is selected as the next switch in the tree.…”

“…Figure 2 shows the new link direction assignment for the same network graph depicted in Figure 1. It has been shown that the DFS methodology [9] provides more minimal paths than the BFS one, resulting in a significant increase in network performance [8]. …”

“…This problem becomes more important as network size increases. In general, up*/down* concentrates traffic near the root switch, often providing minimal paths only between switches that are allocated near the root switch [8], [7]. Additionally, the concentration of traffic in the vicinity of the root switch causes a premature saturation of the network, thus obtaining a low network throughput and leading to an uneven channel utilization.…”