Institute for Systems Research Technical Reports

Permanent URI for this collectionhttp://hdl.handle.net/1903/4376

This archive contains a collection of reports generated by the faculty and students of the Institute for Systems Research (ISR), a permanent, interdisciplinary research unit in the A. James Clark School of Engineering at the University of Maryland. ISR-based projects are conducted through partnerships with industry and government, bringing together faculty and students from multiple academic departments and colleges across the university.

Browse

Filters

19922

Settings

Author: search.filters.author.Nelson, R.Subject: search.filters.subject.Communication

Search Results

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Distributed Parallelism Considered Harmful
    (1992) Makowski, Armand M.; Nelson, R.; ISR
    We consider a model of a distributed parallel processing system that shows that parallel versus sequential processing is beneficial only under conditions of light load. Our results are valid under general assumptions on the number of processors, task service times and the information used to schedule jobs. Our model of a parallel processing system consists of a set of homogeneous processors each with private memory in which tasks queue before being served. Jobs arriving to the system consist of a random number of tasks which can be executed independently each other and we consider a job to be completed only after all of its component tasks have finished execution. a central dispatcher schedules the tasks on the processors at job arrival instants using information on the number of tasks currently scheduled on each processor. We model this system as a distributed fork/join queueing system and derive the structure of the individually optimal scheduling policy. Our results show that the individually optimal is a mixture of policies corresponding to sequential job execution (all tasks are scheduled on a single processor) and parallel scheduling (tasks are distributed among several processors in a manner that tends to equalize queue lengths). We show that, under conditions that include the case of moderate to heavy loads, the individually optimal scheduler schedules tasks according to the sequential policy which runs counter to the intuition that parallel processing is desirable. Because we do not include certain overheads associated with executing jobs in parallel in our model, our results are biased towards parallel rather than sequential processing. since we believe that systems are not typically underutilized, our results strongly suggest that it can be harmful to have parallel execution in distributed processing systems. Response time properties of the individually optimal scheduler are derived and compared to other scheduling policies.
  • Thumbnail Image
    Item
    Optimal Scheduling for a Distributed Parallel Processing Model
    (1992) Makowski, Armand M.; Nelson, R.; ISR
    We consider a model of a parallel processing system consisting of K distributed homogeneous processors each with private memory in which tasks queue before being served. Jobs arriving to the system consist of a set of tasks which can be executed independently of each other and we consider a job to be completed only after all of its component tasks have finished execution. A central dispatcher schedules the tasks on the processors at job arrival instants using information on the number of tasks currently scheduled on each processor. We model this system as a distributed fork/join queueing system and derive the structure of the individually optimal scheduling policy. Our results show that the individually optimal policy is a mixture of policies corresponding to sequential job execution (all tasks are scheduled on a single processor) and parallel scheduling (tasks are distributed among several processors in a manner that tends to equalize queue lengths). We show that, under conditions that include the case of moderate to heavy loads, the individually optimal scheduler schedules tasks according to the sequential policy which runs counter to the intuition that parallel processing is desirable. Because we do not include certain overheads associated with executing jobs in parallel in our model, our results are biased towards parallel rather than sequential processing. Thus our results strongly suggest that for actual distributed memory systems the benefits of parallel processing can be achieved only in conditions of light load. Response time properties of the individually optimal scheduler are derived and compared by simulation to other scheduling policies.