Parallel Job Scheduling under Dynamic Workloads

Abstract

Abstract. Jobs that run on parallel systems that use gang scheduling for multiprogramming may interact with each other in various ways. These interactions are affected by system parameters such as the level of multiprogramming and the scheduling time quantum. A careful evaluation is therefore required in order to find parameter values that lead to optimal performance. We perform a detailed performance evaluation of three factors affecting scheduling systems running dynamic workloads: multiprogramming level, time quantum, and the use of backfilling for queue management -and how they depend on offered load. Our evaluation is based on synthetic MPI applications running on a real cluster that actually implements the various scheduling schemes. Our results demonstrate the importance of both components of the gang-scheduling plus backfilling combination: gang scheduling reduces response time and slowdown, and backfilling allows doing so with a limited multiprogramming level. This is further improved by using flexible coscheduling rather than strict gang scheduling, as this reduces the constraints and allows for a denser packing.

The Dynamics of Backfilling: Solving the Mystery of Why Increased Inaccuracy May Help

“…Based on the information available to the scheduler at T 0 (time 0), it appears the earliest time for J 3 (job 3) to start is T 12 , even though the real earliest start time is actually T 6 . Thus, the scheduler makes a reservation on J 3 's behalf for T 12 and can only backfill jobs that honor this reservation. At T 4 , J 2 terminates.…”

“…The f -model is the dominant model for generating artificial user runtime estimates. It is used to complement workloads that lack estimates data [36,12,13], but more importantly, to evaluate the impact of inaccurate user estimates on backfilling algorithms [30,10,38,36,24,1,28,37,26,5,14]. Based on the f -model, researchers have drawn neat conclusions that range from "performance is independent of accuracy", through "what the scheduler don't know won't hurt it", to "inaccuracy actually improves performance" (see Introduction).…”

“…1 They termed f the "badness factor" because estimates become increasingly inaccurate as f grows, with f = 0 indicating completely accurate estimates. The f -model has been used when simulating workloads that lacked estimates data [36,12,13], but much more importantly, the model and its variants have been extensively used to study the impact of inaccurate estimates on backfilling algorithms [30,10,38,36,24,1,28,37,26,5,14]. One simple variant of interest is the "deterministic f -model", in which there is no randomness and estimates are a direct multiple of the runtime and the badness factor plus one: (f + 1) · r [38,1,5].…”

Time-Sharing Redux for Large-Scale HPC Systems

“…We treat the first 100 days as a warm-up period, to avoid the common warm-up pitfall of job scheduler simulations [16]. Saturation [17] occurs around 90%, which means that variations in utilization are largely driven by job arrival times, rather than scheduling effects [18].…”

Parallel Job Scheduling — A Status Report

“…Improved performance is obtained by using backfilling, and allowing small jobs to move ahead in the queue [53,52]. In fact, using backfilling fully compensates for the loss of performance due to the limited number of jobs that are actually run concurrently [21].…”