Elba Garza will present her MSE thesis talk on May 12, 2015 at 1:30pm in CS 401.
Elba Garza will present her MSE thesis talk on May 12, 2015 at 1:30pm in CS 401. The members of her committee are Margaret Martonosi, Doug Clark, and Kelly Shaw (University of Richmond). Everyone is invited to attend her talk. Her abstract follows below. Title: Efficient Design Space Exploration Techniques in Heterogeneous Systems Abstract: Heterogeneous architectures allow for computational offloading to various compute units such as graphics processing units (GPUs) and application-specific integrated circuits (ASICs). These compute units often require proper hardware and software parameter settings to reach power and performance goals. Optimization of these parameter setting configurations can be conducted through exploration of the parameters’ cumulative design space. Exhaustive exploration is often impractical or impossible due to spaces’ exponential growth with the addition of parameters. For example, a GPU design space analyzed in this work included 47 billion possible parameter set- ting configurations. Utilizing a statistical learning technique called Starchart, we have been able to optimize design space exploration for parameters in some heterogeneous architecture components. We extend Starchart to create Iterative Starchart, which allows non-uniform iterative sampling from certain parts of the design space. Iterative sampling allows the user to focus on areas of interest or importance to the space, whether it be due to performance measurements or necessary parameter constraints. This extension to Starchart results in efficient and tractable design space assessment. To show these properties, we apply Iterative Starchart to re-explore the real-data GPU design spaces from the original publication. We show Iterative Starchart ex- tends prediction performance accuracy by 21% in high-performance areas of interest. Furthermore, we explore design spaces of fixed-function accelerators simulated via, Aladdin, a power-performance simulator. In these accelerator spaces, Iterative Starchart extends prediction performance accuracy by 14% in high-performance areas of interest. These two test cases show Iterative Starchart ca be a valuable and powerful tool for exploring the design spaces of compute units in heterogeneous architectures.
participants (1)
-
Nicki Gotsis