CNS Core: Small: Gmem: Unified Address Space Management for Heterogeneous Systems

点击翻译按钮获取中文摘要

With the decline in year-over-year performance increases achieved by new general-purpose processors, a growing variety of devices, e.g., field-programmable gate arrays, are being added to computers to accelerate applications. Consequently, the operating system (OS) must not only manage access to the computer's memory by applications running on general-purpose processors but also on a diverse array of accelerators. Currently, memory access by these accelerators is managed by different, disconnected parts of the OS, leading to a proliferation of independent, redundant, and often inefficient implementations. This project aims to replace the current fragmented architecture by a unified framework for all devices.Specifically, this project's principal objective is to create a generalized mapping and protection system, called Gmem, that will provide a single, consistent interface within the OS for the creation, management, and destruction of mappings to any physical memory across all devices. Gmem will simplify the overall architecture of the mapping system within the OS. It will enable the authors of device drivers to implement just the device-specific behavior that they need, rather than implementing a complete mapping system. Furthermore, Gmem will provide asynchronous interfaces and employ coupled page table management, which together will enable high performance and strong protection.In its never-ending quest for increased performance, the computing community should expect to see the continued rise of specialized accelerators, increased integration with graphics processing units, networking interfaces, and storage devices, and large amounts of non-volatile memory across the computing system. This conglomeration of components in next generation computing systems requires OS designers to rethink the ways in which the OS manages disparate memory. A broader impact of this project is that it will contribute to the significant innovations in OS software that will be needed to realize the full potential of these systems.Software artifacts and experimental data of long-term value that are created for this project will be available at https://github.com/rice-systems/gmem for at least three years beyond the project's end.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.