Date/Time: Wednesday, August 13, 2008, 10:00 – 12:00
Location: CSRI/90 (NM)

TALK 1
Title: The Los Alamos Roadrunner Hybrid Petascale Supercomputer
Speaker: Ben Bergen, Los Alamos National Laboratory

Brief Abstract Roadrunner is a hybrid supercomputer that will be installed at Los Alamos National Laboratory (LANL) in 2008.  It will be the world’s first general purpose machine to achieve a sustained petaflop.  Roadrunner reaches this level of performance by boosting a conventional Opteron cluster-of-clusters with IBM Cell processors.  Prior to embarking on this new architectural path, LANL commissioned the Roadrunner Advanced Algorithms Team to adapt several important codes to Roadrunner’s projected architecture.  The speedups demonstrated on prototype hardware and software were part of the conclusion that a hybrid, heterogeneous supercomputer is programmable and can support the Laboratory’s national security mission.  This presentation describes the programming models and selected techniques used by the Roadrunner Advanced Algorithms Team in carrying out this work.

This presentation also provides details and highlights from the early access applications runs performed on Roadrunner in Poughkeepsie, New York, during the week of May 9th, 2008.  Three separate codes were successfully run on the full Roadrunner system (17 CUs) resulting in several ground breaking achievements in computational science.  Additionally, several science runs were performed and verified with existing data.  These results show that Roadrunner is a viable platform for meeting the current and future needs of the laboratory stewardship program.  Results are presented for VPIC (plasma physics), SPaSM (molecular dynamics), and PetaVision (neural networks).

TALK 2
Title: Bottom up” reimplementation of the Jayenne Project’s Implicit Monte Carlo for Roadrunner
Speaker: Timothy Kelley, Los Alamos National Laboratory

Brief Abstract  We present one of the two approaches used to reimplement our Implicit Monte Carlo code for hybrid Roadrunner.  We begin with an overview of the mainline code, pointing out aspects that are important in developing a hybrid code.  We then describe the development of the particle-pushing code on the Cell SPE.  We relate the early performance results, and how we changed the code to improve performance.  Mistakes are recounted and lessons learned from them.  We then turn to the process of imbedding the Cell code into a full, hybrid application, with emphasis on design points.  The performance of the full application as measured on prototype hardware for the Roadrunner Phase 3 assessment is presented, as well as subsequently measured performance on actual Roadrunner hardware.