Co-scheduling algorithms for high-throughput workload execution

Abstract

International audienceThis paper investigates co-scheduling algorithms for processing a set of parallel applications. Instead of executing each application one by one, using a maximum degree of parallelism for each of them, we aim at scheduling several applications concurrently. We partition the original application set into a series of packs, which are executed one by one. A pack comprises several applications, each of them with an assigned number of processors, with the constraint that the total number of processors assigned within a pack does not exceed the maximum number of available processors. The objective is to determine a partition into packs, and an assignment of processors to applications, that minimize the sum of the execution times of the packs. We thoroughly study the complexity of this optimization problem , and propose several heuristics that exhibit very good performance on a variety of workloads, whose application execution times model profiles of parallel scientific codes. We show that co-scheduling leads to faster workload completion time (40% improvement on average over traditional scheduling) and to faster response times (50% improvement). Hence co-scheduling increases system throughput and saves energy, leading to significant benefits from both the user and system perspectives

Scheduling Parallel Tasks under Multiple Resources: List Scheduling vs. Pack Scheduling

“…The results extend the theoretical analyses of single-resource-type scheduling [29,2] to account for the presence of multiple resource types.…”

“…The speedup function relates the execution time of a task to the amount of resources allocated to it, and the total area is defined as the product of execution time and resource allocation. Many prior works [2,5,22] have assumed that the speedup of a task is a nondecreasing function of the amount of allocated resources (hence the execution time is a non-increasing function) and that the total area is a non-decreasing function of the allocated resources. One example is the well-known Amdahl's law [1], which specifies the speedup of executing a parallel task with s sequential fraction using p processors as 1/(s + 1−s p ).…”

“…Both algorithms pack rectangles onto shelves, which are equivalent to creating pack-based schedules. The first result (i.e., 3-approximation) has also been extended to the case of moldable task scheduling [2,29]. Turek et al [29] presented a strategy to extend any algorithm for scheduling rigid tasks into an algorithm for scheduling moldable tasks in polynomial time.…”

“…Between these two paradigms, list scheduling can make better use of the available resources (at least in theory) by reducing the idle times between the tasks and thus maximizing the resource utilization. However, pack scheduling has been advocated by some recent studies [2,27] for being practically valuable due to its ease of implementation (as batch processing) and for incurring less scheduling overhead. The relative merits of the two scheduling paradigms are yet to be evaluated under multiple-resource constraints.…”

Resilient co-scheduling of malleable applications

“…This work provides an important extension to our previous work on coschedules [3], which already demonstrated that sharing the platform between two or more applications can lead to significant performance and energy savings [2]. To the best of our knowledge, it is the first work to consider co-schedules and failures, and hence to use malleable applications to allow redistributions of processors between applications.…”