UTCS Colloquium/AI: Greg Hamerly/Baylor University Improving SimPoint Accuracy for Small Simulation Budgets with EDCM Clustering ACES 2.402 Wed. April 30 2008 at 1:00 p.m.

Contact Name: 
Jenna Whitney
Date: 
Apr 30, 2008 1:00pm - 2:00pm

Type of Talk: UTCS Colloquium/AI

Speaker/A

ffiliation: Greg Hamerly/Baylor University

Date/Time: Wednesday A

pril 30 2008 1:00 p.m.

Location: ACES 2.402

Host: Emmett

Witchel

Talk Title: Improving SimPoint Accuracy for Small Simulatio

n Budgets with EDCM Clustering

Talk Abstract:
Detailed processor

simulation is extremely costly on large
benchmark suites where each pr

ogram may run for billions
of instructions and take months of simulatio

n time. We can
obtain good approximate answers in less time using limit

ed
simulation but deciding which regions to simulate is a difficult problem. SimPoint is one approach for choosing simulation
regions ba

sed on the k-means clustering algorithm. We
propose using an alternati

ve clustering model based on a
mixture of exponential Dirichlet compoun

d multinomial (EDCM)
models. This method outperforms k-means in perfor

mance
prediction accuracy when simulation budgets are limited. The

EDCM mixture can cluster high-dimension frequency vector
data directly
without dimension reduction and trains quickly.

Speaker Bio:
Gr

eg Hamerly is an assistant professor of computer science
at Baylor Univ

ersity. His research is in machine learning
particularly in unsupervi

sed learning algorithms and their
applications. He is a primary contri

butor to the SimPoint
project which uses unsupervised learning for eff

icient
computer processor simulation.