SHF: Medium: Type-Specific Instruction Processing

点击翻译按钮获取中文摘要

This research project develops hardware software interfaces to enable standardsoftware programs to benefit from specialized hardware accelerators.Standard CPUs are general purpose meaning that they are designed to executeall types of applications -- from graphics processing in video games tospreadsheets. Because of the diversity of workloads they are expected toexecute, general purpose processors can never be overly specialized for anyindividual one. By contrast, processors which are tailored to a singleworkload or workload domain are narrower in their applicability but moreefficient at executing applications in their target domains.This research project examines specialized processors that target abstractdatatypes: moderate to large scale data structures and their associatedoperations. Modern software engineering practice encourages the use of suchabstract types to improve programmer efficiency and software reliability.Abstract datatype processors thus align with these practices making themapping of software to hardware intuitive and streamlined. By deliveringenergy-efficient, application-specific hardware in an easy-to-program fashion,this research empowers programmers to write faster software that consumes lessenergy.The research activities span three fields of computer science: Hardware systemand architecture research is carried out in software simulation. This portionof the research explores multiple aspects of the hardware system includingefficient implementations of software-style polymorphism and mechanisms toenforce data encapsulation. The project is grounded in a specific,performance-critical, real-world problem of database query processing. Thiscomponent of the research identifies target types for hardware accelerationthat are used in common, complex database operations such as rangepartitioning. Performance results will be obtained both by direct measurementand by simulation. Finally, the compiler segment of the project developscompiler techniques to link high-level languages to the accelerators availableon the target hardware system. The compiler adapts software at runtime tobest utilize the available accelerators and to partition code amonggeneral-purpose and specialized processing cores. As power and energy become the most critical resource in computing systems,mechanisms to conserve these resources have come under significant scientificscrutiny. The broader impact of this research project is to harness softwareabstractions in specialized, energy efficient hardware. By prioritizingprogrammability and efficiency equally, this research strives for usable,energy-efficient, next-generation computing platforms. As a first area ofapplication, this project targets databases, the software engines that storeand index and compute on society's information.