Large-grain pipelining on distributed-memory multiprocessors
The distributed nature of distributed-memory multiprocessors increases the system's extendibility and modularity. However, it also introduces communication difficulties/emdash/data must be moved around explicityly among the processors in the system. In this paper, the concept of largegrain pipelining is introduced, which attempts to maximize the overlapping within the system to minimize the effects of communication overhead. The basic idea is to regulate the information flows in the system in a pipelined fashion. The concept is contrasted with that of systolic arrays, and a general procedure is outlined which transform a systolic algorithm into an algorithm using large-grain piplining. An analytic model ahs been developed, which, not only assists in determining optimal design in a more formal way. Predicted performance from the analytic model matches closely with experimental results obtained from a 64-node NCUBE, and both confirm the effectiveness of largegrain pipelining.
- Research Organization:
- Michigan State Univ., East Lansing (USA). Dept. of Computer Science; Argonne National Lab., IL (USA)
- DOE Contract Number:
- W-31109-ENG-38
- OSTI ID:
- 5114051
- Report Number(s):
- CONF-880567-5; CONF-880567-; ON: DE88009974
- Country of Publication:
- United States
- Language:
- English
Similar Records
Models for parallel computation on distributed-memory multiprocessors
Pipelined data parallel algorithms: Concept and modeling