CAREER: Cross-Layer Schemes For Flexible Resource Sharing in Multicore Systems

Purdue University, West Lafayette IN

Investigators

Abstract

PROJECT ABSTRACT: The proposal addresses the challenge of flexibly and efficiently managing the allocation of shared caches and shared memory bandwidth among applications/threads for diverse sharing scenarios (co-operative, competitive and adversarial) in multicore systems. The proposed solution which spans the hardware, operating system (OS) and application layers, consists of two parts, one at the hardware-OS boundary and another at the OS-application boundary. The first part proposes a hardware-assisted, OS-driven resource allocation technique that effectively combines the flexibility of OS-based resource management with the efficiency of hardware resource management. The second part proposes adoption of currency-based market mechanisms at the application/OS boundary wherein contenders negotiate/bid for access to resources in exchange for some notional currency. Market mechanisms, which have been studied for other resource allocation problems, incentivize good behavior because it is in the participants' interest to make efficient use of resources and release resources that are not used. In the specific context of shared caches and shared memory bandwidth, market mechanisms allow development of sophisticated management policies with prioritization across and within applications. With appropriate interfaces, they also enable several other optimizations at the OS and application layers such as (a) adaptive applications that modify their behavior according to resource availability, (b) co-scheduling of sharers with complementary resource needs and (c) demand-driven pricing" of resources. The key components of the education plan include (a) integration of parallelism and multicore concepts into the graduate and undergraduate computer architecture related courses, (b) development of a graduate-level, advanced computer architecture course to focus on architecture design patterns recurring architecture optimizations similar to the high-level, object-oriented programming design patterns" captured by Gamma et.al. and (c) fostering research participation at both undergraduate and graduate levels.