Last-Level Cache Interference-Aware Scheduling in Scientific Clouds

Cloud infrastructures are increasingly being adopted in different purposes, like business applications and scientific experiments. Relying in virtualization, it addresses several advantages such as resource consolidation and improved management of entire infrastructure, leading to opportunity to reduce costs associated with power consumption and acquisition of new hardware. However, consolidation of multiple applications onto multi-core servers can impose significant performance degradation due to sharing of hardware resources, such as last-level cache. Such phenomenon of performance interference among co-located applications can become problematic, since applications' required quality-of-service (QoS) will not met. In this work, we analyse the cache interference in virtual machines' execution time, and we propose an adapted algorithm to dynamically map virtual machines to physical hosts, so as to regenerate and meet jobs' S LA constraints. Our decision making algorithm detects and reacts to performance degradation, due to last- level cache sharing, and leverages existing virtualization mechanisms to reconfigure the virtual infrastructure , in an accurately and controlled manner. Compared to state of the art approaches the results show that our algorithm performs better targeting users' required performance over the cloud infrastructure.