Models for Parallel Computation

Parallel computing must meet seemingly contradictory goals: software development should be carried out in an architecture-and technology-independent environment, while both algorithms and system software should take full advantage of the features of the underlying parallel architecture. At the same time, application programmers cannot be expected to be parallel processing experts. Usability, scalability, and portability are thus central issues in high performance computing. A recognized important step towards achieving these goals is the development of a computational model which accurately reeects constraints of existing machines , has broad applicability with respect to existing and future machines, and allows accurate prediction of performance. Such a computational model should provide a bridge between software and hardware and fullll the same function as the von Neumann model for sequential computation. In recent years parallel models have been proposed reened, and evaluated. While no model seems to yet Research supported in part by ARPA under contract DABT63-92-C-0022ONR. The views and conclusions contained in this paper are those of the authors and should not be interpreted as representing oocial policies, expressed or implied, of the U.S. government. fullll the stated goals, signiicant progress has been made. This session contains three papers providing diierent viewpoints on parallel models. In the following we give a brief description of the development of parallel models and the papers appearing in this session. The discussion makes references to various models, but the list is not intended to be exhaustive. The PRAM model, a simple shared memory model, was introduced in 9] and it has proven useful in identifying the inherent parallelism in problems. The PRAM has allowed the development of architecture-independent parallel algorithms at a time when parallel platforms were changing rapidly 15, 17, 22]. Some PRAM algorithms have been successfully translated into eecient distributed memory algorithms. For other, especially ones with highly data-dependent communication patterns, eecient adaptations have proven diicult. The PRAM model is widely considered to not match the essential features of existing parallel machines. The main objections include the lack of being able to account for communication la-tency as well as memory and network contention. The paper by Gibbons argues the important role the PRAM model still plays 11]. He describes a new shared memory model, the Queued Shared Memory Model (QSM) which accounts for limited communication bandwidth by evaluating the maximum memory 1 contention in each phase of the algorithm. Gibbons shows that the QSM model has …