Performance variability in advanced computing systems, such as those supporting the cloud computing paradigm, is growing intractably and leads to inefficiency and resource wastage. A key requirement in large-scale virtualized infrastructure, e.g., Amazon EC2, Microsoft Azure, etc., is to provide a guaranteed quality of service to cloud tenants, especially in today's multi-tenant cloud environments. This generally involves using past information and prediction of the probability distribution of requests to match resources that meet service-level agreements. The variability in systems performance hinders the cloud service providers' ability to effectively guarantee SLAs, and thus efficiently meet user demands. In this paper, we propose innovative methodologies for resource management, which leverages the understanding of performance variability in high performance computing systems to exploit new opportunities for tradeoffs between system stability and performance in the cloud. This would help cloud providers better provision and design their infrastructure, as well as ensure meeting provider-tenant SLAs. Moreover, the approach also leads to improved cloud service costs, as tighter bounds on variability could be codified in cost structures bundled in operations or directly offered to cloud tenants.
展开▼
