Savant: Automatic parallelization of a scheduling heuristic with machine learning

Abstract

Abstract-This paper investigates the automatic parallelization of a heuristic for an NP-complete problem, with machine learning. The objective is to automatically design a new concurrent algorithm that finds solutions of comparable quality to the original heuristic. Our approach, called Savant, is inspired from the Savant syndrome. Its concurrency model is based on mapreduce. The approach is evaluated with the well-known Min-Min heuristic. Simulation results on two problem sizes are promising, the produced algorithm is able to find solutions of comparable quality.

Savant: Automatic generation of a parallel scheduling heuristic for map-reduce

“…The accuracy of the mappers is worse than for the smaller problems. The average accuracy increased from the 58% of the original study [9] to 63%. This confirms that the 16 machines to assign tasks to need more training observations than for the 4 machines of the smaller 128 × 4 problems.…”

“…Finally, Pinel et al [33] proposed two parallel designs for GPU of Min-Min and a cGA, reporting speedups of up to 538 times faster with respect to the sequential Min-Min. Additionally, Pinel et al [9] presented a novel framework, inspired in the Savant syndrome, that automatically learns and reproduce the behavior of a target algorithm. The resulting algorithm can be executed using map-reduce framework, so it can be run on large clusters, as well as on multi-core and GPU architectures.…”

“…The concurrent design intentionally focuses on small problems, where the source of concurrency does not lie in the data or the complexity of the algorithm, as suggested by D. Hillis [8]. This paper extends our previous work [9]: the training configuration is refined, the algorithm's capability is improved, and the results are further analyzed to better understand its behavior. Section 2 describes the problem and reviews previous work.…”

A Survey of Evolutionary Computation for Resource Management of Processing in Cloud Computing [Review Article]

“…One of the emerging objectives is the minimization of the environmental impact, driven by the policies of governments and companies, the technical challenges in reliable provisioning of high cloud BrokErinG [27], [28], [29], [30], [ [38], [39], [40], [16], [14], [41], [42], [43] [42] [44] workflow schEdulinG [45], [46], [47], [48], [15], [49], [50] [15] [47], [51], [48], [15] [37] [21] [52], [53] dynaMic caPaciTy PlanninG [54], [18] [17], [18] [55] [19] sErvEr farM load BalancinG [56] [55] [57] [58], [59] [ 56] power, and the cost of electricity for largescale IT systems. Environmental objectives include the minimization of electrical energy use, the peak power draw of the systems, cooling costs, and CO 2 emissions.…”

“…Pinel et al [59] propose another novel method using an algorithm based on the savant syndrome [74] to optimize load balancing on heterogeneous system. The algorithm structure imitates the ways of finding solutions to computationally hard problems employed by autistic people, using a distributed pattern recognition scheme.…”