Architectural and Complier Mechanisms for Accelerating Single Thread Applications on Mulitcore Processors.

Abstract

Multicore systems have become the dominant mainstream computing platform. One of the biggest challenges going forward is how to efficiently utilize the ever increasing computational power provided by multicore systems. Applications with large amounts of explicit thread-level parallelism naturally scale performance with the number of cores. However, single-thread applications realize little to no gains from multicore systems.

This work investigates architectural and compiler mechanisms to automatically accelerate single thread applications on multicore processors by efficiently exploiting three types of parallelism across multiple cores: instruction level parallelism (ILP), fine-grain thread level parallelism (TLP), and speculative loop level parallelism (LLP).

A multicore architecture called Voltron is proposed to exploit different types of parallelism. Voltron can organize the cores for execution in either coupled or decoupled mode. In coupled mode, several in-order cores are coalesced to emulate a wide-issue VLIW processor. In decoupled mode, the cores execute a set of fine-grain communicating threads extracted by the compiler. By executing fine-grain threads in parallel, Voltron provides coarse-grained out-of-order execution capability using in-order cores. Architectural mechanisms for speculative execution of loop iterations are also supported under the decoupled mode. Voltron can dynamically switch between two modes with low overhead to exploit the best form of available parallelism.

This dissertation also investigates compiler techniques to exploit different types of parallelism on the proposed architecture. First, this work proposes compiler techniques to manage multiple instruction streams to collectively function as a single logical stream on a conventional VLIW to exploit ILP. Second, this work studies compiler algorithms to extract fine-grain threads. Third, this dissertation proposes a series of systematic compiler transformations and a general code generation framework to expose hidden speculative LLP hindered by register and memory dependences in the code. These transformations collectively remove inter-iteration dependences that are caused by subsets of isolatable instructions, are unwindable, or occur infrequently.

Experimental results show that proposed mechanisms can achieve speedups of 1.33 and 1.14 on 4 core machines by exploiting ILP and TLP respectively. The proposed transformations increase the DOALL loop coverage in applications from 27% to 61%, resulting in a speedup of 1.84 on 4 core systems.