Page 171 of the Proceedings

Distributed Preemptive Scheduling on Windows NT Donald McLaughlin and Partha Dasgupta Arizona State University partha@asu.edu This research is partially supported by grants from DARPA/Rome Labs, Intel Corporation and NSF.

Introduction	All multitasking operating systems use preemptive scheduling. Many multiprocessor systems also employ preemptive inter-task scheduling when they run parallel computations. However, preemptive scheduling in distributed systems is rare, if not non-existent. Consider a cluster of workstations, running a parallel application. The application divides itself into a set of tasks. The scheduler assigns these tasks to a set of workstations. Often the tasks are not of equal length, the machines are not of equal speeds and tasks can create further subtasks. These situations lead to non-optimal matches of workers to tasks causing executions that do not complete as quickly as it would be possible in a better matched case. Also, the granularities of the tasks may be small, leading to high overhead.
Distributed Scheduling	All multitasking operating systems use preemptive scheduling. Many multiprocessor systems also employ preemptive inter-task scheduling when they run parallel computations. However, preemptive scheduling in distributed systems is rare, if not non-existent. Consider a cluster of workstations, running a parallel application. The application divides itself into a set of tasks. The scheduler assigns these tasks to a set of workstations. Often the tasks are not of equal length, the machines are not of equal speeds and tasks can create further subtasks. These situations lead to non-optimal matches of workers to tasks causing executions that do not complete as quickly as it would be possible in a better matched case. Also, the granularities of the tasks may be small, leading to high overhead.
Preemptive Scheduling	All multitasking operating systems use preemptive scheduling. Many multiprocessor systems also employ preemptive inter-task scheduling when they run parallel computations. However, preemptive scheduling in distributed systems is rare, if not non-existent. Consider a cluster of workstations, running a parallel application. The application divides itself into a set of tasks. The scheduler assigns these tasks to a set of workstations. Often the tasks are not of equal length, the machines are not of equal speeds and tasks can create further subtasks. These situations lead to non-optimal matches of workers to tasks causing executions that do not complete as quickly as it would be possible in a better matched case. Also, the granularities of the tasks may be small, leading to high overhead.
Implementation and Performance	We have implemented the algorithms on the Chime parallel processing system running on Windows NT. The major roadblock turned out to be process migration under NT. The lack of signals posed the greatest problem as a thread can only be interrupted by another thread that suspends it. Care has to be taken to ensure that the thread to be migrated is not suspended waiting for a runtime event. Race conditions and starvation conditions have been encountered. The final system runs well, and performance results are very encouraging. We found that the round-robin scheduler provided acceptable performance on large grained programs, but was hampered by the migration overhead. The task bunching scheduler performed really well in a wide variety of situations. More information and papers can be found at https://milan.eas.asu.edu.