Exploiting reconfigurable heterogenous parallel architectures in a multitasking context: a systems approach

Abstract

In the recent years, the continuous performance increase of the sequential execution on single CPU systems is facing an upper bound because it primarily relied on the respective circuit operating frequency improvement, which has already reached its limits. This low-level performance bottleneck caused a chain reaction to the above abstraction layers and practically changed the way computing systems are being built. Nowadays, realizing parallelism at all the design levels of a computing platform is the main goal of the respective domain research efforts. Application developers need to improve their skills and take into account architecture-level platform details to decide the optimal application partitioning, while respective re-targetable toolchains have been developed to automate tasks and abstract complexity to the extent possible. On the other hand, traditional general-purpose operating system support and related concepts have not been particularly reconsidered in the emerging broader context of parallel applications, tools and architectures but primarily focus on the efficient task scheduling on shared memory homogeneous and symmetric multicore systems of limited scalability. In this dissertation, we introduce new concepts at the operating system-level to take advantage of the runtime reconfiguration of hardware to exploit its benefits under a general purpose context. Regarding application development, we demonstrate the effectiveness of two different application development framework approaches that are tightly integrated with novel operating system support for optimal execution on many-element arrays. We outline how these contributions allow the operating system to efficiently distribute any type of platform resources, deal with performance asymmetry and load balance, at runtime, parallel applications that execute concurrently on the emerging parallel reconfigurable platforms. We have implemented prototypes of every proposed concept and we report the results of real life experiments.