hmats

HMATS (Hybrid Machine–Assignment and Task Scheduler) is a software framework designed to automate the allocation of computing resources to heterogeneous workloads in distributed computing environments. It was first introduced by the University of Edinburgh’s Cloud Systems Lab in 2018 as a response to the growing complexity of resource management in multi‑tenant cloud infrastructures. The core idea behind HMATS is to combine rule‑based policies with machine‑learning models to predict task performance and resource requirements, thereby improving utilization and reducing execution time.

The architecture of HMATS is divided into three main components: the Policy Engine, the Performance Predictor, and the Scheduler. The Policy Engine enforces administrative constraints such as priority levels, cost limits, and fairness quotas. The Performance Predictor uses historical execution traces and real‑time metrics to forecast CPU, memory, and I/O demands for new tasks. Finally, the Scheduler integrates the outputs of the Policy Engine and the Performance Predictor to issue placement decisions to the underlying resource manager, such as Kubernetes or OpenStack. HMATS supports live migration of containers and virtual machines, enabling dynamic load balancing without downtime.

Early deployments of HMATS were carried out at several research data centers, where it achieved a 15‑20 % reduction in average job completion time compared with traditional static schedulers. Commercial interest emerged in 2021, and a version based on HMATS was incorporated into the open‑source project OpenCloudScheduler. Critics argue that the machine‑learning component can introduce unpredictability and may create opacity in decision making. Nevertheless, HMATS has been cited in multiple peer‑reviewed studies that examine advanced resource allocation strategies for mixed‑workload cloud environments.