Torta: Extending applications' capability to select heterogeneous computing resources adaptively to the PC usage

Spread of GPU-accelerated applications on PCs can cause serious degradation of the user experience such as frame dropping on the video playback, due to applications' resource competition on the same GPU due to arbitrary processors selection. In this paper, we propose a processors assignment system for real applications that achieves processors assignment according to condition based rules without modifying applications. To demonstrate the feasibility of our concept, we implemented a prototype of the centralized processors assignment mechanism called Torta. Our experiment using eight practical applications has shown that Torta achieves binary-compatible processors switching with an average performance penalty on only 0.2%. In a particular case where a video playback application is executed with three other GPU-intensive applications, our method enables users to enjoy the video playback with 60 frames per second (FPS) while the FPS decreases to 14 without the mechanism. This paper shows the design and the implementation of Torta on Windows 7 and concludes that our mechanism increases the efficiency of computational resource usage on PCs, thus improves the overall user experiences.