Disaggregated memory architectures for blade servers

Current trends in memory capacity and power of servers indicate the need for memory system redesign. Memory capacity is projected to grow at a smaller rate relative to the growth in compute capacity, leading to a potential memory capacity wall in future systems. Furthermore, per-server memory demands are increasing due to large-memory applications, virtual machine consolidation, and bigger operating system footprints. The large amount of memory required is leading to memory power being a substantial and growing portion of server power budgets. As these capacity and power trends continue, a new memory architecture is needed that provides increased capacity and maximizes resource efficiency.;This thesis presents the design of a disaggregated memory architecture for blade servers that provides expanded memory capacity and dynamic capacity sharing across multiple servers. Unlike traditional architectures that co-locate compute and memory resources, the proposed design disaggregates a portion of the servers' memory, which is then assembled in separate memory blades optimized for both capacity and power usage. The servers access memory blades through a redesigned memory hierarchy that is extended to include a remote level that augments local memory. Through the shared interconnect of blade enclosures, multiple compute blades can connect to a single memory blade and dynamically share its capacity. This sharing increases resource efficiency by taking advantage of the differing memory utilization patterns of the compute blades.;This thesis evaluates two system architectures that provide operating system-transparent access to the memory blade; one uses virtualization and a commodity-based interconnect, and the other uses minor hardware additions and a high-speed interconnect. The ability to extend and share memory can achieve orders of magnitude performance improvements in cases where applications run out of memory capacity, and similar improvements in performance-per-dollar in cases where systems are overprovisioned for peak memory usage. To complement the evaluation, a hypervisor-based prototype of one system architecture is developed. Finally, by extending the principles of disaggregation to both compute and memory resources, new server architectures are proposed for large-scale data centers that can double performance-per-dollar when considering total cost of ownership compared to traditional servers.