University of Illinois
Department of Electrical and Computer Engineering
ECE 411 Fall 2003
Sanjay J. Patel, Instructor
Assignment #3
Due October 15th
For this assignment you are required to experiment with branch prediction. Using the
high level description of branch predictors described in class (the prediction point is 100s
of instructions ahead of the execution point, which is 10s of instructions ahead of the
retirement point), devise some interesting investigations on the correlation between
various phenomenon (e.g., previous branches, subroutine calls, data values, etc) and
branches. This assignment is vague by design because I want to stimulate you to try
something creative in hopes of discovering phenomenon that might lead to improved
predictor design. If you are struggling for ideas, please come and talk to me.
To seed your imagination, here are some ideas:
• Fourier or entropy analysis of branch behavior
• Redo the Evers et al study on different benchmarks. Do the results still hold?
• Apply machine learning techniques to predictor design
• Discover a type of predictor that predicts a particular class of branches really well.
• Examine predictor scaling: With more transistors, do standard global predictors
continue to become more accurate. That is, do we really need more innovation in
predictor design?
To assist you in this assignment, I have provided a stripped down version of the
SimpleScalar instruction-level simulator and SPEC2000 benchmark suite. You are
welcome to use these if you wish. Or you can go with your own too.
Please write up your findings in a 2 to 3 page paper. Include diagrams, graphs, and
references to assist the reader in comprehending your findings.
On Oct 16th, we will have very short presentations from randomly selected students.
