A Survey on Thread-Level Speculation Techniques

Alvaro Estebanez,Diego R Llanos,Arturo Gonzalez-Escribano

doi:10.1145/2938369

Abstract

Thread-Level Speculation (TLS) is a promising technique that allows the parallel execution of sequential code without relying on a prior, compile-time-dependence analysis. In this work, we introduce the technique, present a taxonomy of TLS solutions, and summarize and put into perspective the most relevant advances in this field.

Highlights

Thread-Level Speculation (TLS), called Speculative Parallelization (SP), or even Optimistic Parallelization, is a runtime technique that executes in parallel fragments of code that were originally intended to run sequentially
In [Torrellas 2011], an accurate summary of Thread-Level Speculation techniques is given, including a detailed description of the two main issues that any TLS system should solve: How to buffer and manage speculative states, and how to detect and handle dependence violations. His analysis makes any effort to reproduce a summary of TLS characteristics here meaningless: we suggest the reader to consult his work to better understand the fundamentals of the field and the management of side effects due to the use of thread-level speculation
The main application of TLS is in the parallel execution of irregular doacross loops when the total number of dependences that appear at runtime is low

Summary

Introduction

Thread-Level Speculation (TLS), called Speculative Parallelization (SP), or even Optimistic Parallelization, is a runtime technique that executes in parallel fragments of code that were originally intended to run sequentially. Speculative threads should detect whether they have consumed a datum that was subsequently updated by a predeccessor thread, that is, a thread executing an earlier part of the code, according to sequential semantics. Such situations, called dependence violations, should be detected and rectified by hardware or software mechanisms, or a combination of both, to keep sequential semantics.

Objectives

Methods

Findings

Conclusion