Scheduling Data Transfers in Parallel Computers and Communications Systems

Ravi Jain


PhD Dissertation
Dept of Computer Sciences
Univ of Texas at Austin
December 1992

Also available as Tech Report TR-93-03
Advisors: J. C. Browne and John S. Werth
250 pp.


ABSTRACT

The performance of many applications of parallel computers and communications systems is limited by the speed of data transfers rather than the speed of processing.  An important, but neglected, aspect of resource management to overcome this bottleneck is the scheduling of data transfers.  Data transfer scheduling differs from traditional scheduling problems in that data transfer tasks require multiple resources simultaneously, rather than a single resource serially, in order to execute.

We study the data transfer scheduling problem by first defining a general model for precisely specifying and classifying scheduling problems.  We use the model for the recognition of the similarity of seemingly different problems from different application areas, for the systematic transformation of one problem specification into that of a seemingly different problem, and for the systematic decomposition of a problem specification into solvable subproblems.

We obtain polynomial-time, optimal and approximate algorithms for a wide range of data transfer scheduling problems under a variety of architectural and logical constraints, including communication architectures in which resources are fully connected, communication architectures with a tree topology, and the presence of mutual exclusion and precedence constraints.  Our algorithms either generalize previous results for these problems, or provide better performance, or both.

Our results are applicable to both parallel computers and communications systems, including certain types of shared-bus multiprocessor systems such as the Sequent and the IBM RP3, hierarchical switching systems, tree-structured multiprocessor architectures, and intersatellite communications systems.