A synthesis algorithm for reconfigurable interconnection networks

Abstract

The performance of a parallel algorithm depends in part on the interconnection topology of the target parallel system. An interconnection network is called reconfigurable if its topology can be changed between different algorithm executions. Since communication patterns vary from one parallel algorithm to another, a reconfigurable network can effectively support algorithms with different communication requirements. In this paper, we describe how to generate a network topology that is optimized with respect to the communication patterns of a given task. The algorithm presented takes as input a task graph and generates as output a topology that closely matches the given input graph. The topologies generated by our algorithm are analyzed with respect to optimum interconnection topologies for the best, worst, and average cases. Simulation results verify the average case performance prediction and confirm that, on the average, the optimum topologies are generated.Index Terms-Degree constrained connected graph, degree constrained subgraph, mapping algorithm, reconfigurable interconnection network, synthesis algorithm.OO18-9340/88/06OO-0691$01 .OO 0 1988 IEEE Insup Lee (S'82-M'82-S'82-M'83) received the B.

Synthesizing minimum total expansion topologies for reconfigurable interconnection networks

“…As was shown using the cardinality synthesis algorithm [6], c is almost always (probability approachs 1 as the number of nodes in the graph for the graphs used in the 24 and 100 node simulation. For the r=3 plots z was set to 1.13.…”

“…The 0 FIE (G) I) complexity of factoring combined with the 0 (n IE (P) I ) complexity of the hill climbing part of the algorithm limited the maximum size to approximately 100 nodes (approximately 45 minutes of CPU time for a 100 node graph on a HF'9000-300 workstation). For the degree bound, we used r=3, 6 and To determine performance with respect to the density or number of edges in the task graph, a wide range of edge probabilities was used starting at ri(n-1). This starting probability insured that the average degree of a node in the graph was greater than or equal to r. Tasks graphs in which all nodes have degree less than r have trivial optimal solutions.…”

“…We note that there is some debate as to what it means for a network topology to match the communication requirements of an algorithm [6]. In this paper, a parallel algorithm is represented by a task graph.…”

Reconfigurable interconnection networks in Distributed Shared Memory systems: a study on communication patterns

“…Developing parallel algorithms whose communication patterns closely match a certain interconnection topology of a target parallel system is often a difficult and non-convenient task. And although one interconnection topology may have the ideal node distribution for data interchange for a set of algorithms, it may not be well suited to other set, introducing large latencies on the network for other algorithms [1].…”

“…The reconfigurable interconnection networks for multiprocessor systems have attracted a lot of research interests [1][2][3][4][5]. An interconnection network is called reconfigurable if its topology can be changed during program execution, usually to match the communication requirements of an application.…”