UCLA

Resource disaggregation has emerged as a promising solution to enhance both resource utilization and management efficiency in datacenters. Existing disaggregation solutions have largely centered on generic, low-level system optimizations such as minimizing remote access latency at the operating system and hardware levels. However, these solutions often yield suboptimal performance due to the lack of alignment between application semantics and the underlying system layers.

This dissertation presents a novel approach that enhances the performance of disaggregated systems by incorporating application semantics, including memory access patterns, data object ownership, and computational intensity, into the system design. Our methodology is demonstrated through three techniques -- Mako, MemLiner, and DRust. Each technique applies program semantics at different levels of the system stack, ranging from programming languages and compilers to runtime environments and operating systems. Specifically, Mako and MemLiner utilize program semantics to develop a new runtime that is optimized for disaggregated memory architectures. Meanwhile, DRust employs data object ownership semantics in applications to build a programming framework tailored for compute disaggregation.

Collectively, these proposed techniques aim to enhance the performance, efficiency, and consistency of disaggregated systems, making them more viable for practical implementation in today's datacenters. This body of work lays a foundational framework for the co-design and co-optimization of techniques across system layers, aimed at advancing future disaggregated datacenters.