High Performance Computing as a Service with Service Level Agreements

Cloud computing has been gaining popularity for quite some time in various areas, on the infrastructure, platform and application level. Recently, the possibility to provide high performance computing (HPC) as a service has been investigated in conjunction with the cloud computing paradigm. While this is a viable solution for applications that do not require HPC in the truest sense -- with supercomputers which offer paramount performance regarding computation, network interconnect and storage -- there are HPC applications which cannot be realized in this way. HPC as a service can be offered for these applications as well, but it requires a different approach than the usage of cloud computing. The enhancement of mostly best effort based HPC with long-term service level agreements (SLAs) is a potential solution. HPC providers then need not only to decide on which service levels to offer but need to closely investigate the framework conditions for these service levels. The scheduling of these service levels is a difficult task and we simulate a proposed algorithm for providing guarantees on waiting times and the implications on other jobs. We investigate the influence of setting a maximum allowed job size on prioritized jobs and conclude that it makes sense to restrict this size for both clients and provider.