Automated Problem Mapping: the Crystal Runtime System.
Abstract
The effective utilization of multiprocessors, particularly those with architectures that cannot support shared memory in an efficient way, is currently dependent on the ability of the user to map the problem onto the multiprocessor. In order to obtain high levels of efficiency, this mapping must distribute computational load relatively evenly between the machine's processors and must minimize the effects of interprocessor communication delay on algorithm performance. The need to explicitly designate a problem decomposition and to verify that the decomposition is both correct and has the desired performance characteristics can be a time consuming and error prone task. In cases in which the load distribution of and algorithm cannot be predicted sufficiently well in advance to allow a deterministic decomposition to be specified, it may be necessary to specify a family of problem decompositions along with a procedure for run time load management. While the development of methods for dynamically balancing loads is an active area of current research without the development of automated mapping methods, the implementation of such schemes can be particularly time consuming and difficult. A methodology will be developed that can insulate the user from the considerations required to produce efficient programs for multiprocessor machines while still enabling the user to achieve high levels of performance. Programs will be written in a very high level programming language Crystal. Keywords: Parallel processing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1987
- Accession Number
- ADA177563
Entities
People
- Joel H. Saltz
- Marina C. Chen
Organizations
- Yale University