Using service level agreements in a high-performance computing environment

Abstract

The concept of Service Level Agreements (SLAs) has come to attention particularly in conjunction with Grid computing. SLAs allow for a controlled collaboration between partners, that is service providers and their customers. SLAs have originated in the telecommunication industry but have found broad uptake in the Grid computing community, especially regarding the topics of their automated negotiation, general management and legal implications. High-performance computing (HPC) providers, however, have not been taking up SLAs yet, even though they often provide not only supercomputer resources but at the same time access to cluster or grid resources. This might be due to the fact that SLAs are mostly regarded on a per-job basis, which is not consistent with the contractual model that HPC providers currently use. In this work, we analyze how HPC providers can implement SLAs and what benefits this brings, proposing the usage of SLAs on a long-term basis. We further present a software that has been developed to simulate the scheduling of jobs at an HPC provider site using service levels and analyze how different distributions of service levels between the submitted jobs influence machine usage and average waiting times of jobs. Finally, we present how an HPC provider can practically implement SLA-based scheduling.