Scientific Computing Seminar

Date:

Thursday, March 2, 2006

Time:

1:00pm-2:00pm

Location:

70-191

Seminar Speaker:

Olaf Lubeck
Los Alamos National Laboratory

Title:

A Monte Carlo Model of In-order Micro-architectural Performance: Decomposing Processor Stalls

Abstract:

Modeling has long been the key to gaining insight and understanding of complex systems. For many years, we have successfully modeled the application efficiency of parallel clusters given a measurement of execution time of an application on a single processor node. In this work, we propose a predictive model for a single processor node of an in-order micro-architecture based on Monte Carlo methods that is validated against the Itanium 2. The model uses statistical application characteristics to predict CPI with an average error of 7%. Since prediction is achieved in a few seconds, the model can be used to gain insight into the exploration space that has been the domain of cycle-accurate simulation. The work can provide a basis for performance analysis of future micro-architectures that are pushing the envelope in multi-core, SIMD and multi-threaded features, as well as provide a foundation for application characterization.

Sponsor of Seminar:

John Shalf

Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov