Response time and parallelism in parallel-processing systems with certain synchronization constraints
What makes the exact analysis of parallel-processing systems so problematic is the internal parallelism within jobs, namely, a job may need and consequently may hold more than one processor at a time. A multiprocessor system is viewed as a set of P cooperating processors, and a computer job as a set of tasks partially ordered by some precedence relationships. For a finite number of processors, it is assumed that the total-system capacity is shared among the jobs proportionate to the number of their ready tasks. This is non-egalitarian processor sharing, as compared with the usual egalitarian processor-sharing technique. Two fundamental performance measures of concern in a multiprocessor system are the expected job sojourn time and the achievable system speedup. The author constructs models and methodologies to analyze these two measures by exploiting the underlying stochastic processes. In those cases where the exact analysis fails, he provides bounds and/or approximate solutions backed up by simulations of the exact models. Specifically, he investigates the probability distribution of the number of occupied processors, the generating function of this distribution, and its first two moments. In particular, the expected number of busy processors is found to be dependent only on the average number of tasks per job, the job average arrival rate, and the task average processing requirement.
- Research Organization:
- California Univ., Los Angeles (USA)
- OSTI ID:
- 7055340
- Country of Publication:
- United States
- Language:
- English
Similar Records
Scheduling parallel programs in distributed systems
A parallel Lauritzen-Spiegelhalter algorithm for probabilistic inference