Data Sharing and Resource Management in the Multicore Era
Speaker(s) : Sandhya Dwarkadas (University of Rochester)
Achieving high resource utilization and efficiency is hampered on modern multicore-based systems by their lack of performance transparency. Increasing compute and storage capabilities, often heterogeneous in their properties, have resulted in performance challenges due to the non-uniformity of compute and data access costs, as well as contention for shared resources. In this talk, I will present our efforts to improve the scalability and efficiency of data sharing and resource management for parallel and multiprogrammed workloads. I will describe our use of hardware performance counter information to track sharing behavior and control resource allocation within the operating system. I will also outline our design of memory management metadata to share address translation information and eliminate redundancy.