ASA - The Adaptive Scheduling Architecture

Abstract

In High Performance Computing (HPC) infrastructures, resources are controlled by batch systems and may not be readily available, which can negatively impact applications with deadlines and long queue waiting times. In particular, this is noticeable for data intensive and low latency workflows where resource planning and timely allocation are key characteristics for efficient processing. On the one hand, allocating the maximum capacity expected for a scientific workflow guarantees the fastest possible execution time, at the cost of spare and idle infrastructural resources, as well as extended queue waiting times and costly resource usage. On the other hand, dynamically allocating resources according to specific workflow stage requirements optimizes resource usage, although it may also negatively impact the total workflow makespan. With the aim of enabling new scheduling strategies and features for scientific workflows, we propose ASA: the Adaptive Scheduling Architecture, a novel and convergence proven scheduling method to reduce perceived queue waiting times as well as to optimize resource usage and planning in scientific workflows. The algorithm uses reinforcement learning to estimate queue waiting times, and based on these estimates pro-actively submits resource change requests, with the goal of minimizing total workflow inter-stage waiting times, idle resources, and makespan. The algorithm takes into consideration both learning (the waiting times), and acts on what is learnt so far, and thus handles the exploration-exploitation trade-off. Experiments with real scientific workflows in two real supercomputers show that ASA combines the best of the two aforementioned approaches for resource allocation, with average workflows' queue waiting time and makespan reductions of up to 10% and 2% respectively, with up to 100% prediction accuracy, while obtaining near optimal resource utilization.

A HPC Co-scheduler with Reinforcement Learning

“…ASA -The Adaptive Scheduling Architecture is an architecture and RL algorithm that reduces the user-perceived waiting times by accurately estimating queue waiting times, as well as optimizes scientific workflows resource and cluster usage [36]. ASA encapsulates application processes into containers and enable fine-grained control of resources through and across job allocations.…”

“…In contrast to that type of ML, RL algorithms such as the one used in ASA X and in [43] do not require historical training data to optimize scheduling. As mentioned in previous sections, ASA X is a stateful extension of [36], where the main difference is that ASA X can incorporate previous decisions in a RL approach. In [39], the authors combine offline job classification with online RL to improve collocations.…”