An automatically partitioning compiler for SISAL
In this paper, we describe a compiler that automatically repackages SISAL programs to achieve the grain size needed to run efficiently on different multiprocessors. The compiler is based on an existing implementation for SISAL on the Sequent Balance multiprocessor, and on previous work on automatic partitioning of SISAL programs. The granularity of parallelism in the existing Sequent implementation is defined by language constructs (e.g., all function calls are spawned as separate tasks), causing the programming style to dramatically effect multiprocessor performance. It is desirable for the partitioning to be performed automatically, so that the same program can be made to execute efficiently on different multiprocessors and the programmer need not be concerned with granularity issues when developing his program. This is the goal of our automatically partitioning compiler. The partitioner operates on IF1, a graphical intermediate language for SISAL programs. It also takes as input a list of parameters describing the target multiprocessor. The partitioner's output consists of IF1 annotations to specify which calls to user-defined functions should be executed as parallel tasks, which Forall expressions should be executed as parallel loops and their appropriate chunk sizes. the Sequent Balance implementation has been retargetted to the Alliant, multi-VAX and Cray X-MP multiprocessor systems. It is currently being modified to support the partitioner's output, rather than just using the default granularity. We intend to use the automatically partitioning compiler to study the performance of various SISAL application programs on these different multiprocessors. 11 refs., 8 tabs.
- Research Organization:
- Lawrence Livermore National Lab., CA (USA)
- DOE Contract Number:
- W-7405-ENG-48
- OSTI ID:
- 6432700
- Report Number(s):
- UCRL-98289; CONF-880975-2; ON: DE89006738
- Country of Publication:
- United States
- Language:
- English
Similar Records
POSC: A Partitioning and Optimizing SISAL Compiler
Implementation and performance of a domain decomposition algorithm in Sisal