The increasing heterogeneity of computing systems enables higher performance and power efficiency. However, these improvements come at the cost of increasing the overall complexity of designing such systems. These complexities include constructing implementations for various types of processors, setting up and configuring communication protocols, and efficiently scheduling the computational work. The process for developing such systems is iterative and time consuming, with no well-defined performance goal. Current performance estimation approaches use source code implementations that require experienced developers and time to produce.
We present a framework to aid in the design of heterogeneous systems and the performance tuning of applications. Our framework supports system construction: integrating custom hardware accelerators with existing cores into processors, integrating processors into cohesive systems, and mapping computations to processors to achieve overall application performance and efficient hardware usage. It also facilitates effective design space exploration using processor models (for both existing and future processors) that do not require source code implementations to estimate performance.
We evaluate our framework using a variety of applications and implement them in systems ranging from low power embedded systems-on-chip (SoC) to high performance systems consisting of commercial-off-the-shelf (COTS) components. We show how the design process is improved, reducing the number of design iterations and unnecessary source code development ultimately leading to higher performing efficient systems.
Library of Congress Subject Headings
Field programmable gate arrays--Design and construction; Graphics processing uings
Computing and Information Sciences (Ph.D.)
Sonia Lopez Alarcon
Skalicky, Samuel, "A Model-based Design Framework for Application-specific Heterogeneous Systems" (2015). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus