Job Scheduling for Loosely-Coupled Inhomogeneous Nodes Using Data Envelopment Analysis

Abstract

Job Scheduling in high performance computing (HPC) clusters and grids has traditionally been performed by job entry and management sys tems, such as the Portable Batch System that place their emphasis on job management and only to a lesser extent on job scheduling. In grid infrastruc tures and emerging, virtual machine-based HPC environments, the previous assumption on relative homogeneity of nodes does not hold any more. In con trast, loosely coupled nodes in these settings are more heterogenous than ev er. This places new demands on job scheduling, where a large number of dif ferent nodes create the problem of optimally laying out compute jobs across the network for efficient resource allocation. The proposed approach present ed utilizes non-parametric Data Envelopment Analysis (DEA) to derive a workload-type proximity factor for a given node type. An experimental fac tor determination is performed using 5 physical and one virtual nodes.