CSR: Small: Big Memory: Exploring Memory Management Mechanisms and Policies for Rack-Scale Computers

William Marsh Rice University, Houston TX

Investigators

Abstract

Large-memory applications, including data analytics, in-memory object caches, and in-memory databases, are a major consumer of data center resources. Moreover, these applications can effectively use more memory than a single computer can provide. Traditional approaches to scaling these applications involve parallelism across multiple computers. However, that may not use resources cost-effectively, as the ratio of processing, networking, and memory is determined by the configuration of the underlying computers, not the application. This project explores disaggregated servers, or "rack-scale computers". These servers consist of pools of processors, memory, and I/O devices within a rack that can be flexibly allocated to virtual machines based on the needs of their resident applications. This project focuses on one aspect of disaggregated servers: managing a large pool of memory within the rack. Specifically, using this pool to dynamically add memory to and remove memory from virtual machines. This will enable applications to handle bursty load without the need to spin up/down entire virtual machines. In order to effectively manage memory across virtual machines, the hypervisor will need a better understanding of how memory is being used, by both the guest operating systems and applications. Therefore, we are exploring the interfaces, mechanisms, and policies within and between the hypervisor, guest operating system, and application that will enable the use of large disaggregated memories. This research is one step towards rethinking what a computer is and how to construct one that better serves data science and analytics applications that can benefit from large amounts of memory.