CSR: Small: Predictable Real-Time Computing in GPU-enabled Systems

点击翻译按钮获取中文摘要

Given the need to achieve higher performance without driving up energy consumption, most chip manufacturers have shifted to multicore architectures, especially heterogeneous ones. Among heterogeneous processing elements, graphic processing units (GPUs) have seen wide-spread use. GPUs have the power to enable orders of magnitude faster execution of many applications. Thus, they are becoming increasingly applicable for general-purpose systems. Unfortunately, it is not straightforward to reliably adopt GPUs in many safety-critical systems that require predictable real-time correctness, one of the most important tenets in certification required for such systems. A key example is the advanced automotive system where timeliness of computations is an essential requirement of correctness due to the interaction with the physical world. The goal of this project is to ensure predictable real-time correctness in current GPU-enabled systems, through (i) developing new real-time resource allocation methods that can be applied in GPU-enabled systems, where a number of difficult analysis issues due to the problem of co-scheduling CPU and GPU resources and several GPU-specific constraints will be addressed, and (ii) building an open-source ecosystem of predictably managing GPU resources in the operating system. These efforts will pave the way to utilizing GPUs in a predictable manner and benefit many applications and systems in which real-time constraints exist, such as automated automobiles and medical instrumentation. The successful completion of this work will enable powerful GPU computing capability. In safety-critical systems which usually require certification, this work will help enable such systems equipped with GPUs to be certifiable.