Agenda: The slides are provided for each presentation. For each predictor, the writeup and code is provided. The code is provided as a uuencoded bzipped tar file (.tar.bz2.uue) formatted according to the directions found by clicking here. The traces are not provided with the code, but are provided separately as part of the CBP framework. The CBP framework can be downloaded from the following page. None of the predictors require the traces with data values and memory addresses, so you don't need to download those traces.
1:00 Introduction: Competition Overview and Workshop Agenda, Jared Stark, MRL/MTL Intel
1:10 Perceptrons for Dummies, Daniel A. Jiménez, Rutgers University
1:30 Idealized Piecewise Linear Branch Prediction, Daniel A. Jiménez, Rutgers University
1:50 A PPM-like, Tag-based Predictor, Pierre Michaud, IRISA/INRIA
2:10 Adaptive Information Processing: An Effective Way to Improve Perceptron Predictors
Hongliang Gao and Huiyang Zhou, University of Central Florida
2:30 Break
3:00 A 2bcgskew Predictor Fused by a Redundant History Skewed Perceptron Predictor
Veerle Desmet, Hans Vandierendonck, and Koen De Bosschere, Ghent University
3:20 The O-GEHL Branch Predictor, André Seznec, IRISA/INRIA
3:40 The Frankenpredictor, Gabriel Loh, Georgia Institute of Technology
4:00 Branch Prediction Caveats and Second-Order Effects, Phil Emma, IBM Research
4:20 Conclusion: Ranking of the Finalists, Anointing of the Champion, and “What Next?”
Chris Wilkerson, MRL/MTL Intel
4:30 Adjourn
Results: The following two tables show the average mispredict rates, in mispredicts per 1000 instructions, for the 6 finalists. The finalists are indicated by the first name of the first author on the predictor. For reference, the mispredict rates of an equivalently-sized gshare predictor are also given. All CBP contestants were scored using the average mispredict rate across all traces (ALL column in the tables). The average mispredict rates across the floating-point (FP), integer (INT), multi-media (MM), and server (SERV) traces are also given for reference.
The first table is for the undistributed trace set, which, as its name implies, was not distributed to the finalists, and was used in the final round of the competition to rank the finalists and crown the champion. The second table is for the distributed trace set, which was distributed to all contestants, and was used in the initial round of the competition to rank the contestants and choose the finalists.
|
|
FP |
INT |
MM |
SERV |
ALL |
|---|---|---|---|---|---|
|
Hongliang |
0.483 |
2.917 |
3.590 |
3.304 |
2.574 |
|
André |
0.567 |
3.467 |
3.627 |
2.845 |
2.627 |
|
Gabe |
0.596 |
3.369 |
3.695 |
3.140 |
2.700 |
|
Daniel |
0.629 |
2.963 |
3.884 |
3.492 |
2.742 |
|
Pierre |
0.659 |
3.691 |
3.854 |
2.905 |
2.777 |
|
Veerle |
0.617 |
3.210 |
3.670 |
3.732 |
2.807 |
|
GSHARE |
1.553 |
4.691 |
5.484 |
6.351 |
4.520 |
|
|
FP |
INT |
MM |
SERV |
ALL |
|---|---|---|---|---|---|
|
André |
0.590 |
3.299 |
4.450 |
2.939 |
2.820 |
|
Hongliang |
0.428 |
3.223 |
4.444 |
3.194 |
2.823 |
|
Daniel |
0.678 |
3.237 |
4.567 |
3.327 |
2.952 |
|
Gabe |
0.601 |
3.836 |
4.727 |
3.181 |
3.086 |
|
Pierre |
0.672 |
3.811 |
4.934 |
2.989 |
3.101 |
|
Veerle |
0.681 |
3.698 |
4.529 |
3.789 |
3.174 |
|
GSHARE |
1.301 |
6.607 |
6.778 |
6.520 |
5.301 |