Computational & Technology Resources
an online resource for computational,
engineering & technology publications
PROCEEDINGS OF THE SECOND INTERNATIONAL CONFERENCE ON PARALLEL, DISTRIBUTED, GRID AND CLOUD COMPUTING FOR ENGINEERING
Edited by: P. Iványi and B.H.V. Topping
Providing QoS through Service Level Agreements in High Performance Computing
High Performance Computing Center Stuttgart, Stuttgart, Germany
R. Kübert, "Providing QoS through Service Level Agreements in High Performance Computing", in P. Iványi, B.H.V. Topping, (Editors), "Proceedings of the Second International Conference on Parallel, Distributed, Grid and Cloud Computing for Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 52, 2011. doi:10.4203/ccp.95.52
Keywords: service level agreements, high-performance computing, cloud computing, quality of service, scheduling.
Instead, users submit their compute jobs through dedicated machines. What has not changed, however, is the fact that there are no service levels available to users. At best, different job sizes are treated differently, apart from that, jobs are mostly treated in a best-effort, first come first served way, with some notion of fairness applied that is completely defined by the service provider. Even the advent of service level agreements (SLAs), which allow a simple specification of quality of service (QoS) guarantees, has not led to changes for HPC providers: if at all, then trivial QoS guarantees are supported and SLAs have to be negotiated on a per-job basis, which means there is a potential for negotiation overhead and the risk of not reaching an agreement.
This paper presents a different approach using SLAs as long-term contracts where jobs are submitted in reference to a previously established contract. It is shown how different service levels can be set up at an HPC provider's site and what this entails for both provider and customer. A workload trace has been used to analyse the impact of providing service levels on the different service classes and the results were compared with a classical scheduling without service levels. A service provider may keep machine usage unchanged, even after introducing different service levels with different priorities. The influence on the average waiting time of jobs depends on their service level; average waiting time over all jobs regardless of their service class is increased. The influence of the provision of service levels on the two factors that have been analysed leads us to the conclusion that different key objective functions have to be investigated in the future: previously, mostly machine usage and adherence to deadlines has been investigated while other objectives have not been regarded. However, when service levels are provided, the adherence of the provider to the guarantees given for the service levels has to take precedence and only then can other objective functions be considered.
purchase the full-text of this paper (price £20)