Date of Award:

2016

Document Type:

Thesis

Degree Name:

Master of Science (MS)

Department:

Electrical and Computer Engineering

Advisor/Chair:

Sanghamitra Roy

Abstract

Timing speculation is a promising approach to increase the processor performance and energy efficiency. Under timing speculation, an integrated circuit is allowed to operate at a speed faster than its slowest path|the critical path. It is based on the empirical observation, which is presented later in the thesis, that these critical path delays are rarely manifested during the program execution. Consequently, as long as the processor is equipped with an error detection and recovery mechanism, its performance can be increased and/or energy consumption reduced beyond that achievable by any other conventional operation.

While many past works have dealt with timing speculation within a single core, in this work, a new direction is being uncovered | timing speculation for a multi-core processor executing a parallel, multi-threaded application. Through a rigorous cross-layered circuit architectural analysis, it is observed that during the execution of a multi-threaded program, there is a significant variation in circuit delay characteristics across different threads.

Synergistic Timing Speculation (SynTS) is proposed to exploit this variation (heterogeneity) in path sensitization delays, to jointly optimize the energy and execution time of the many-core processor. In particular, SynTS uses a sampling based online error probability estimation technique, coupled with a polynomial time algorithm, to optimally determine the voltage, frequency and the amount of timing speculation for each thread. The experimental analysis is presented for three pipe stages, namely, Decode, SimpleALU and ComplexALU, with a reduction in Energy Delay Product by up to 26%, 25% and 7.5% respectively, compared to existing per-core timing speculation scheme. The analysis also embeds a case study for a General Purpose Graphics Processing Unit.

Checksum

823953debccb02b06c7148648d2017c4

Share

COinS