Low-Cost Approximation Algorithms for Scheduling Independent Tasks on Hybrid Platforms

Hybrid platforms embedding accelerators such as GPUs or Xeon Phis are increasingly used in computing. When scheduling tasks on such platforms, one has to take into account that a task execution time depends on the type of core used to execute it. We focus on the problem of minimizing the total completion time (or makespan) when scheduling independent tasks on two processor types, also known as the \((Pm,Pk)||C_{\max }\) problem. We propose BalancedEstimate and BalancedMakespan, two novel 2-approximation algorithms with low complexity. Their approximation ratio is both on par with the best approximation algorithms using dual approximation techniques (which are, thus, of high complexity) and significantly smaller than the approximation ratio of existing low-cost approximation algorithms. We compared both algorithms by simulations to existing strategies in different scenarios. These simulations showed that their performance is among the best ones in all cases.