ROSES: an efficient scheduler for precedence - constrained tasks on concurrent multiprocessors
One of the first steps in attempting to solve a problem on a concurrent computation ensemble is its decomposition into a set of tasks (or processes). Problem decomposition often induces precedence constraints among the tasks, which the distributed nature of the computational system translates into message passing requirements. The efficient scheduling of sets of tasks among which complex interrelationships (such as precedence constraints) may exist is essential for achieving optimal performance from the new generation of concurrent supercomputers. This paper focuses on the ROSES system, which is being developed to schedule precedence-constrained tasks for computation by an ensemble of concurrent processors. This endeavor is particularly difficult when the number of tasks required exceeds the number of available processors, of when the interconnection topology of the task graph differs from the interconnection topology of the computation ensemble. The ROSES approach seeks near-optimal solutions by combining heuristic techniques to minimize scheduling time as a function of the number and relationship of tasks, with sophisticated data structures to most efficiently utilize available computer memory, and algorithms to control search and eliminate dead ends. The methodology is applied to the solution of the inverse dynamics equations of a robot manipulator and benchmark results are included. 51 refs., 1 fig., 5 tabs.
- Research Organization:
- Oak Ridge National Lab., TN (USA)
- DOE Contract Number:
- AC05-84OR21400
- OSTI ID:
- 5841358
- Report Number(s):
- CONF-8508173-2; ON: DE86011083
- Country of Publication:
- United States
- Language:
- English
Similar Records
Multiprocessor scheduling with practical constraints
Scheduling with stochastic gert precedence constraints