"The Predictability of Computations that Produce Unpredictable Outcomes"
Tor Aamodt, University of Toronto
Abstract
--------
We study the dynamic stream of slices that lead to branches that foil an
existing branch predictor and to loads that miss and measure whether these
slices exhibit locality (i.e. repetition). We argue that this regularity can be
used to dynamically extract slices for an operation-based predictor that
speculatively pre-computes a load address or branch target (i.e. an outcome)
rather than directly predicting the outcome based upon the history of outcomes.
We study programs from the SPEC2000 suite and find they exhibit good
slicelocality for these problem loads and branches. Moreover, we study the
performance of an idealized operation-based predictor (it can execute slices
instantaneously). We find that it interacts favorably with an existing
sophisticated outcome-based branch predictor, and that slice-locality provides
good insight into the fraction of all branch mispredictions it can potentially
eliminate. Similar observations hold for operation-based prefetching of loads
that miss.
This talk is preparation for a presentation I will give at the 5th Workshop on
Multi-threaded Execution, Architecture, and Compilation, (being held in
conjunction with MICRO-34). Therefore, it will be fairly brief (i.e. 20 minutes
for the talk, plus extra time for Q&A).
Bio
---
Tor Aamodt is a Ph.D. candidate in the Edward S. Rogers Sr. Department of
Electrical and Computer Engineering at the University of Toronto where he also
received his Bachelors degree in Engineering Science and Masters degree in
Computer Engineering. His research interests are high performance computer
architecture and compiler/architecture codesign for embedded computing
applications such as network processing and DSP.
--
Cider Seminars Admin