Job scheduling for acceleration systems in cloud computing

With the increase of various of applications, CPU is no longer adequate for the computation tasks. Accordingly, some providers deploy accelerators in their cloud. Since not all the servers in the cloud can carry accelerators, how to schedule jobs onto accelerators and improve the system performance is an important issue. Due to the distributed computing frameworks in cloud computing systems, the jobs usually arrive in batches, and hence we try to minimize the make-span of a batch of jobs in this paper. To this end, we first formulate this problem as a mathematic programming model, and prove the NP-hardness of this problem. To solve this problem efficiently, we propose a 4- approximation algorithm. Through extensive simulations, we find that our algorithm can reduce the make-span of a batch of jobs by about 32%, and enhance the system throughput by up to 29% compared with our comparison baseline.