Resource Management in Heterogeneous Parallel Computing Environments with Soft and Hard Deadlines

Heuristic-based approaches are often used to perform the assignment and scheduling of parallel high-performance computing applications on machines (as it is an NP-hard problem). The timevarying importance of tasks has been represented using monotonically-decreasing value functions. The value function of a task gives the value that will be earned based on when the task completes execution and is a flexible method to incorporate soft and hard deadlines. To measure the amount of useful work accomplished in an oversubscribed heterogeneous parallel computing environment, we conduct a simulation study to evaluate the performance of schedulers based on the total value they earn from executing tasks. To optimize for such an objective (as opposed to traditional response timebased or fairness-based objectives), requires the design of novel parallel scheduling metaheuristics. We design new metaheuristics and create a new concept of “place-holder” tasks that gives our metaheuristics the advantages of reservations, but also allows future tasks to nullify those reservations. We examine the performance of our metaheuristics and other popular parallel scheduling techniques such as EASY Backfilling and Conservative Backfilling in a variety of environmental conditions. Our real workload trace-driven simulations show that our metaheuristics that use the concept of place-holders consistently outperform the other techniques, in terms of the total value earned, improving on average EASY Backfilling by over 50% and Conservative Backfilling by over 100%.