Search

International audience; In this paper, we propose READYS, a reinforcement learning algorithm for the dynamic scheduling of computations modeled as a Directed Acyclic Graph (DAGs). Our goal is to develop a scheduling algorithm in which allocation and scheduling decisions are made at runtime, based on the state of the system, as performed in runtime systems such as StarPU or ParSEC. Reinforcement Learning is a natural candidate to achieve this task, since its general principle is to build step by step a strategy that, given the state of the system (the state of the resources and a view of the ready tasks and their successors in our case), makes a decision to optimize a global criterion. Moreover, the use of Reinforcement Learning is natural in a context where the duration of tasks (and communications) is stochastic. We propose READYS that combines Graph Convolutional Networks (GCN) with an Actor-Critic Algorithm (A2C): it builds an adaptive representation of the scheduling problem on the fly and learns a scheduling strategy, aiming at minimizing the makespan. A crucial point is that READYS builds a general scheduling strategy which is neither limited to only one specific application or task graph nor one particular problem size, and that can be used to schedule any DAG. We focus on different types of task graphs originating from linear algebra factorization kernels (CHOLESKY, LU, QR) and we consider heterogeneous platforms made of a few CPUs and GPUs. We first propose to analyze the performance of READYS when learning is performed on a given (platform, kernel, problem size) combination. Using simulations, we show that the scheduling agent obtains performances very similar or even superior to algorithms from the literature, and that it is especially powerful when the scheduling environment contains a lot of uncertainty. We additionally demonstrate that our agent exhibits very promising generalization capabilities. To the best of our knowledge, this is the first paper which shows that reinforcement learning can really be used for dynamic DAG scheduling on heterogeneous resources.

Due to the highly irregular nature and prohibitive execution times of Branch-and-Bound (B&B) algorithms applied to combinatorial optimization problems (COPs), their parallelization has received the se two last decades great attention. Indeed, significant efforts have been made to revisit the parallelization of B&B following the rapid evolution of high-performance computing technologies dealing with their associated scientific and technical challenges. However, these parallelization efforts have always been guided by the performance objective setting aside programming productivity. Nevertheless, this latter is crucial for designing ultra-scale algorithms able to harness modern supercomputers which are increasingly complex, including millions of processing cores and heterogeneous building-block devices. In this paper, we investigate the partitioned global address space (PGAS)-based approach using Chapel for the productivity-aware design and implementation of distributed B&B for solving large COPs. The proposed algorithms are intensively experimented using the Flow-shop scheduling problem as a test-case. The Chapel-based implementation is compared to its MPI+X-based traditionally used counterpart in terms of performance, scalability, and productivity. The results show that Chapel is much more expressive and up to 7 . 8 × more productive than MPI+Pthreads. In addition, the Chapel-based search presents performance equivalent to MPI+Pthreads for its best results on 1024 cores and reaches up to 84 % of the linear speedup. However, there are cases where the built-in load balancing provided by Chapel cannot produce regular load among computer nodes. In such cases, the MPI-based search can be up to 4 . 2 × faster and reaches speedups up to 3 × higher than its Chapel-based counterpart. Thorough feedback on the experience is given, pointing out the strengths and limitations of the two opposite approaches (Chapel vs. MPI+X). To the best of our knowledge, the present study is pioneering within the context of exact parallel optimization.