XPS: FULL:CCA: Extracting Scalable Parallelism by Relaxing the Contracts across the System Stack

点击翻译按钮获取中文摘要

Technology scaling trends have made parallelism the de-facto standard for enhancing performance across a spectrum of computing environments spanning from high-end computing to embedded platforms. Yet, the software is woefully lagging in its ability to extract usable parallelism offered by the underlying hardware platforms primarily because of the compartmentalized contracts between the different layers of the system stack. Rigid contracts restrict the ability to leverage a rich design space of performance/power/correctness trade-offs within and across layers, that could be achievable by straying slightly from the contract. Although such a relaxed contract, referred to as approximate computing, has received attention recently, much of the work in this area is still compartmentalized and lacks a holistic cross-layer strategy to maximize parallelism, while adhering to power and correctness mandates. Thus, the motivation of this project is to explore a holistic cross-layer approach to approximate computing spanning application, runtime system, compiler and hardware, thereby breaking the rigidity of the contracts between the layers, while still allowing them to cooperate for extracting the achievable parallelism across a diverse set of applications in both the high-end and mobile computing environments. Specifically, it involves application-level analysis of the scope of approximation for computation, data access and synchronization, designing efficient hardware mechanisms that could facilitate and benefit from approximation, and developing compiler and runtime support for expressing, exploiting and evaluating/validating the approximations in an architecture-aware fashion. This cross-layer approach to approximate computing is expected to play a crucial role towards achieving scalable parallelism for the next decade and beyond, with a potentially high impact to the computing industry. In addition, the tools and models developed from this project are disseminated in the public domain to a broader research community, and the PIs engage in a variety of outreach activities such as recruiting women and minority and involvement of local high school students through Penn State Eberly College's Exploration-U initiatives.