Inductive Bias-driven Reinforcement Learning For Efficient Schedules in Heterogeneous Clusters

The problem of scheduling of workloads onto heterogeneous processors (e.g., CPUs, GPUs, FPGAs) is of fundamental importance in modern datacenters. Most current approaches rely on building application/system-specific heuristics that have to be reinvented on a case-by-case basis. This can be prohibitively expensive and is untenable going forward. In this paper, we propose a domain-driven reinforcement learning (RL) model for scheduling that can be broadly applied to a large class of heterogeneous processors. The key novelty of our approach is (i) the RL model; and (ii) the significant reduction of training-data (using domain knowledge) and -time (using sampling based end-to-end gradient propagation). We demonstrate the approach using real world GPU and FPGA accelerated applications to produce scheduling policies that significantly outperform hand-tuned heuristics.