A Multilayer Approach Towards Optimizing Spatial and Temporal Computing Systems

Abstract

Optimizing computing systems towards efficiency and performance is a fundamental problem of computing systems. This problem continues to be challenging, as new and emerging computing systems exhibit complex performance characteristics influenced by the spatial and temporal details of execution placement and schedule. The execution schedule and placement of a computation is defined and shaped at many levels, potentially including the application structure, a framework, a runtime, the operating system, and the hardware architecture. Thus, it is necessary to optimize for spatial and temporal characteristics throughout a computing stack. 

This dissertation presents works that optimize select computing systems using abstractions that manage spatial and temporal characteristics. Each work addresses a different level in the computing stack. The first work, positioned at the intersection between applications and frameworks, proposes an improvement to state management in serverless systems to improve spatial locality at targeted temporal intervals. The second work, positioned at the intersection between runtimes and operating systems, provides fine-grained, temporally-dynamic resource limits for containers. The third work, and primary effort, includes the design and implementation of a library that provides temporally-dynamic spatial replication of data structures. This library, LD-NR, is positioned to support both applications and operating systems. The third work also includes the design and implementation of a distributed operating system targeting emerging extended non-uniform memory access architectures. The fourth work, positioned at the hardware level, extends and refines a programming interface for a type of neural processing units (NPUs) with an explicitly spatial and temporal design.