EBSCO Logo
Connecting you to content on EBSCOhost
Title

Speculative Precomputation: Exploring the Use of Multithreading for Latency.

Authors

Hong Wang; Wang, Perry H.; Weldon, Ross Dave; Ettinger, Scott M.; Saito, Hideki; Girkar, Milind; Shih-wei Liao, Steve; Shen, John P.

Abstract

Speculative Precomputation (SP) is a technique to improve the latency of single-threaded applications by utilizing idle multi-threading hardware resources to perform aggressive long-range data prefetches. Instead of trying to explicitly parallelize a single-threaded application, SP does the following: • Targets only a small set of static load instructions, called delinquent loads, which incur the most performance degrading cache miss penalties. • Identifies the dependent instruction slice leading to each delinquent load. • Dynamically spawns the slice on a spare hardware thread to speculatively precompute the load address and perform data prefetch. Consequently, a significant amount of cache misses can be overlapped with useful work, thus hiding the memory latency from the critical path in the original program. Fundamentally, contrary to conventional wisdom that multithreading microarchitecture techniques can be used to only improve the throughput of multitasking workloads or the performance of multithreaded programs, SP demonstrates the potential to leverage multithreading hardware resources to exploit a form of implicit thread-level parallelism and significantly speed up single-threaded applications. Most desktop applications in the traditional PC environment are not otherwise easily parallelized to take advantage of multithreading resources. This paper chronicles the milestones and key lessons from Intel's research on SP, including an initial simulation-based evaluation of SP for both in-order and out-of-order multithreaded microarchitectures. We also look at recent experiments in applying software-based SP (SSP) to significantly speed up a set of pointer-intensive applications on a pre-production version of Intel® Xeon™ processors with Hyper-Threading Technology.

Subjects

THREADS (Computer programs); CACHE memory; COMPUTER architecture

Publication

Intel Technology Journal, 2002, Vol 6, Issue 1, p1

ISSN

1535-864X

Publication type

Academic Journal

EBSCO Connect | Privacy policy | Terms of use | Copyright | Manage my cookies
Journals | Subjects | Sitemap
© 2025 EBSCO Industries, Inc. All rights reserved