Joys of Scheduling on Multicore Processors

Past

On multicore processors, processing cores share hardware resources, such as caches, memory controllers and interconnects, and threads running simultaneously on different cores affect each other's performance. A thread scheduler can have a major impact on performance. In particular, if certain threads compete for a certain type of shared resource (e.g., a memory controller), the scheduler can place them such that they do share this resource and run more efficiently as a result. In this talk, I will describe our research on mitigating contention for shared resources via scheduling. This is a challenging problem, because it is difficult to determine at runtime whether or not threads will hurt each other's performance. One school of thought believed that in order to predict the extent of contention between a group of threads, the scheduler must know the threads' memory reuse profiles. But these profiles are very difficult to obtain online. We discovered that contention can be approximated using a much simpler online metric. A contention-aware scheduler built as a result of our research improves upon the worst-case performance of the default Linux scheduler by as much as a factor of two for some applications. I will conclude the talk by discussing some of the on-going research in our group, including contention management on NUMA systems. BIO: Alexandra Fedorova is an Assistant Professor of Computer Science at Simon Fraser University (SFU) in Vancouver, Canada. She earned her Ph.D. at Harvard University in 2006, where she completed a thesis on Operating System Scheduling for Chip Multithreaded Processors. Concurrently with her doctorate studies, Fedorova worked at Sun Microsystems Research Labs, where she investigated transactional memory and operating systems. She is the lead inventor on three US patents, with a few others under review. At SFU, Dr. Fedorova has co-founded Systems, Networking and Architecture (SYNAR) research lab. Presently her research interests span operating systems and virtualization platforms for multi-core processors, with a specific focus on scheduling. Recently she started collaborating on datacenter power efficiency with several local companies. She has more than 20 publications in various academic venues. Her work is supported by Sun Microsystems, Google, Intel and Electronic Arts. She also holds two Strategic Grants from the Canadian government and a BCIC Natural Resources and Applied Sciences grant.