Eliminating on-chip traffic waste: are we there yet?

Smolinski, Robert

Permalink

https://hdl.handle.net/2142/44133 Copy

Description

Title

Eliminating on-chip traffic waste: are we there yet?

Author(s)

Smolinski, Robert

Issue Date

2013-05-24T21:51:45Z

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

Adve, Sarita V.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Multicore cache architecture
• Coherence protocols
• Network traffic
• Energy efficiency

Abstract

As technology continues to scale, the memory hierarchy in processors is predicted to be a major component of the overall system energy budget. This has led many researchers into focusing on techniques that minimize the amount of data moved, and the distance that it is moved. While many techniques have been shown to be successful at reducing the amount of on-chip network traffic, no studies have shown how close a combined approach would come to eliminating all unnecessary data traffic, nor have any studies provided insight into where the remaining challenges are. To answer these questions, this thesis systematically analyzes the on-chip traffic for six applications. For this study, we use an alternative hardware-software coherence protocol, DeNovo, and apply several optimizations to DeNovo to quantitatively show the traffic inefficiencies of a directory-based MESI protocol. With a fully optimized DeNovo protocol, we can remove most of the traffic inefficiencies caused by poor spatial locality, fetch-on-write write policy, poor L2 reuse, and MESI protocol overheads. In our discussion of these improvements, we highlight the software causes of these traffic overheads in order to generalize the results. The final DeNovo protocol with all optimizations applied is able to reduce network traffic by an average of 39.5% and execution time by an average of 10.5%, relative to a baseline MESI implementation. On average, 8.8% of the remaining traffic is spent fetching non-useful data. Using only our optimizations, reducing non-useful data movement further would not be possible without losing performance because of irregular access patterns in the applications.

Graduation Semester

2013-05

Permalink

http://hdl.handle.net/2142/44133

Copyright and License Information

Copyright 2013 Robert Smolinski

Owning Collections