Finding near-perfect parameters for hardware and code optimizations with automatic multi-objective design space explorations

摘要

In the design process of computer systems or processor architectures, typically many different parameters arernexposed to configure, tune, and optimize every component of a system. For evaluations and before production,rnit is desirable to know the best setting for all parameters. Processing speed is no longer the only objectivernthat needs to be optimized; power consumption, area, and so on have become very important. Thus, thernbest configurations have to be found in respect to multiple objectives. In this article, we use a multi-objectiverndesign space exploration tool called Framework for Automatic Design Space Exploration (FADSE) to automaticallyrnfind near-optimal configurations in the vast design space of a processor architecture together withrna tool for code optimizations and hence evaluate both automatically. As example, we use the Grid ALU Processorrn(GAP) and its postlink optimizer called GAPtimize, which can apply feedback-directed and platformspecificrncode optimizations. Our results show that FADSE is able to cope with both design spaces. Less thanrn25% of the maximal reasonable hardware effort for the scalable elements of the GAP is enough to achieve thernprocessor’s performance maximum.With a performance reduction tolerance of 10%, the necessary hardwarerncomplexity can be further reduced by about two-thirds. The found high-quality configurations are analyzed,rnexhibiting strong relationships between the parameters of the GAP, the distribution of complexity, and therntotal performance. These performance numbers can be improved by applying code optimizations concurrentlyrnto optimizing the hardware parameters. FADSE can find near-optimal configurations by effectivelyrncombining and selecting parameters for hardware and code optimizations in a short time. The maximumrnobserved speedup is 15%. With the use of code optimizations, the maximum possible reduction of thernhardware resources, while sustaining the same performance level, is 50%.