Modeling Data Center Co-Tenancy Performance Interference

A multi-core machine allows executing several applications simultaneously. Those jobs are scheduled on different cores and compete for shared resources such as the last level cache and memory bandwidth. Such competitions might cause performance degradation. Data centers often utilize virtualization to provide a certain level of performance isolation. However, some of the shared resources cannot be divided, even in a virtualized system, to ensure complete isolation. If the performance degradation of co-tenancy is not known to the cloud administrator, a data center often has to dedicate a whole machine for a latency-sensitive application to guarantee its quality of service. Co-run scheduling attempts to make good utilization of resources by scheduling compatible jobs into one machine while maintaining their service level agreements. An ideal co-run scheduling scheme requires accurate contention modeling. Recent studies for co-run modeling and scheduling have made steady progress to predict performance for two co-run applications sharing a specific system. This thesis advances co-tenancy modeling in three aspects. First, with an accurate co-run modeling for one system at hand, we propose a regression model to transfer the knowledge and create a model for a new system with different hardware configuration. Second, by examining those programs that yield high prediction errors, we further leverage clustering techniques to create a model for each group of applications that show similar behavior. Clustering helps improve the prediction accuracy of those pathological cases. Third, existing research is typically focused on modeling two application co-run cases. We extend a two-core model to a three- and four-core model by introducing a light-weight micro-kernel that emulates a complicated benchmark through program instrumentation. Our experimental evaluation shows that our cross-architecture model achieves an average prediction error less than 2% for pairwise co-runs across the SPECCPU2006 benchmark suite. For more than two application co-tenancy modeling, we show that our model is more scalable and can achieve an average prediction error of 2-3%