Re-examining dual path processing

Venkit, Abishek

Permalink

https://hdl.handle.net/2142/115330 Copy

Description

Title

Re-examining dual path processing

Author(s)

Venkit, Abishek

Issue Date

2022-01-28

Director of Research (if dissertation) or Advisor (if thesis)

Kumar, Rakesh

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• branch prediction
• computer architecture
• dual path instruction processing
• gem5

Abstract

Branch prediction has long been a heavily studied topic in computer architecture research. Modern branch predictors can achieve upwards of 98% prediction accuracy on many workloads. However, as CPU pipelines have become deeper to support higher clock frequencies, the branch misprediction penalty has increased greatly. Even with high-fidelity branch predictors, branch misprediction accounts for significant performance degradation. To combat this, many have proposed various forms of dual path processing. Dual path processing fetches (and potentially decodes, renames, and executes) instructions on two or more paths when a low-confidence branch is encountered. If the branch is mispredicted, the instructions on the alternate path can immediately be processed, reducing the misprediction penalty. This work re-evaluates the efficacy of dual path processing when paired with modern branch prediction, state-of-the-art confidence estimation, and a deep CPU pipeline. Power and area overheads are considered when designing an architecture to process instructions on multiple paths. We call this architecture Dual Front End (DuFE). Careful analysis of DuFE, branch predictors, and confidence estimators reveal that achieving significant performance gain via dual path processing at a low hardware cost is likely futile. With current confidence estimation schemes, the best-case scenario DuFE variant achieves a 2.87% performance gain with significant overheads.

Graduation Semester

2022-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Abishek Venkit

Owning Collections