Scheduling on Unspecified Heterogeneous Distributed Resources

Abstract

In this paper, we present an adaptive method for scheduling parallel applications on unspecified distributed memory platforms. The presented method can be used to schedule parallel applications when the total workload and the execution parameters (communication speed, available computing power...) are unspecified. When used to schedule divisible load applications according to a masterworker model, this method delivers the workload through multiple rounds. In order to maximize the throughput of each worker, it can prevent both idleness in the use of workers and contentions in the use of the links between master and workers. Before focusing on the proposed scheduling method, the paper recalls the underlying methods on which its development relies. The paper then gives a theoretical analysis of the method before presenting results of simulations obtained with the SimGrid framework on a limited distributed memory platform.

First Experimental Assessments of the Adaptivity of the Scheduling with AS4DR

“…Details about related works can be found in [1] and the proof of the results presented in the previous section has been established in [2]. To the best of our knowledge, no scheduler aims at maximizing the efficiency of the use of the CPUs in a similar context.…”

On constructing DAG‐schedules with large areas

“…A similar delay-of-commitment strategy forms the response advocated in [2,17] to the unpredictability of highly volatile computing platforms. Yet other studies propose scheduling strategies wherein a precomputed static schedule is reorganized at run time in response to changes in processors' powers [36,23]; one instance of this appears in [26], where planned checkpoints allow one to react dynamically to unexpected behavior by a volatile platform.…”

Setting up Clusters of Computing Units to Process Several Data Streams Efficiently

“…Based on these remarks, the AS4DR (Adaptive Scheduling for Distributed Resources) method [15] automatically adapts the schedule to both: the heterogeneity of the workers and the poorness of the estimation of the execution parameters. But, it deals with only one data stream, processed by only one master-workers platform.…”

“…The asymptotic periodicity of the schedule with AS4DR remains when the costs for computations and communications are affine. With the aid of the estimate of the execution parameters and, before the launch of the AS4DR scheduler, a step called CIP (for Contentions and Idleness Prevention) determines a value for τ , large enough to involve without contention all the workers into the scheduling [15]. It has been proved that AS4DR maximizes the CPU-efficiency of the workers in a heterogeneous but steady-state context [16].…”