The high performance solution of irregular problems

Progress in networking technologies now allows the implementation of a 1,000 processor machine which scales linearly in performance The key to realising the potential of these machines is in the support for high performance code which is both scalable (in problem size and number of processors) and portable, using structured software design methods. Work at Leeds has been studying the use of a typed shared memory system, based around the use of Shared Abstract Data Types (SADTs), to provide a structured view of sharing in a parallel machine. This has concentrated on the support for irregular problems, in which the patterns of communication and synchronisation can vary dynamically. SADTs augment the serial ADT with the ability to support concurrent access, encouraging modular and reusable code. The performance of an SADT can be optimised by supporting weak data consistency semantics (where this does not effect the correctness of the given application), and by careful tuning on the given parallel platform. The use of a well defined computational model is critical to the support of scalable and portable SADTs by capturing the dominant machine characteristics which effect performance. The Weakly coherent PRAM (WPRAM) model is an extension of the BSP approach, which has been used to design the SADTs, and to predict their performance. The paper provides a case study to illustrate this approach, based on the travelling salesman problem. Results for a 256 processor Cray T3D machine are provided, which demonstrate the high and scalable performance which has been achieved.