MEMORYISLANDS: A FEDERATED APPROACH FOR EFFICIENT MEMORY MAPPINGS

Abstract

Modern computing systems are exhibiting increasing computing elements with several co-running workloads. These workloads exhibit highly diverse memory access patterns and have different memory requirements. Nonetheless, main memory architectures are still oblivious to this diversity, handling all requests with the same set of rules. Memory mapping is a clear example of this failing "one-size-fits-all" memory approach. Encompassing several parallelism levels (channels, ranks, groups, and banks), the memory performance of an application depends heavily on its particular memory access pattern and how it is mapped to these levels. In contrast, current memory controllers (MCs) deploy a fixed address mapping for all applications, which leaves significant performance opportunities if each application is serviced with the suited mapping. Instead of following the prior approach of attempting to dynamically change the address mapping, which has significant limitations due to the need for data migrations, this thesis promotes the idea of considering main memory as an independent federated set of resources, which we call islands. Based on this idea, it introduces 1) a methodology to decide the address mapping that maximizes the performance of each application; 2) an optimization framework to statically define this federation of islands for each set of co-running workloads; and 3) finally, a software-aware co-design methodology to configure the MC with the various memory islands and their corresponding address mappings. Our extensive evaluation with a diverse set of more than 80 workloads and several single- and multi-core system setups show a significant performance improvement over the best compared static mapping when deploying the proposed technique.