Minimizing memory access cost in embedded systems.

摘要

Embedded systems are extensively utilized in various aspects of the modern world, which fosters great demands on better performance, lower energy consumption, and smaller area. Many researchers have pointed out that caches and memory systems of the embedded applications not only consume the majority of the power, but also become the bottleneck of the performance and area improvement. In order to tackle this problem, some optimization methods that are specially designed for the memory systems on the embedded applications have been proposed, since the embedded applications are often working under the specific environment and the inputs are limited. This dissertation studies the strategy and techniques of minimizing the memory access cost of various embedded systems, including the single-core system, multi-core system, and loop-centric applications.;A hardware-controlled cache, though popular for general computers, may not be the best architectural solution for embedded systems because it consumes large die area and too much energy. Scratch Pad Memories, also known as SPMs, can be used as a substitution of traditional cache for the embedded systems. The difference between SPMs and the traditional caches is that SPMs are controlled by software or programmers. Therefore, it is crucial to design some novel data placement algorithms in order to efficiently utilize the SPMs on an embedded system. Especially, when there are multiple types of memory in a single system, the data placement method always directly affects the performance and the energy cost. In this dissertation, we present a thorough study for multi-level memory systems and propose several optimizing techniques.;On the other hand, for the multi-core embedded systems, traditional data placement methods cannot be simply applied. Therefore, this dissertation proposes a data placement method with data duplication in order to minimize the memory access cost for the whole system. To take advantage of the noticeable pattern of some loop-centric applications, this dissertation also proposes a novel data placement method that can tremendously improve the performance as well as save energy cost.