Login
Computer Science @ Rochester
Monday, September 15, 2003
11:00 AM
CSB 209
Jeffrey S. Vetter
Lawrence Livermore National Laboratory (LLNL)
Performance Modeling and Analysis at LLNL: Striking a Balance between Resolution and Insight
Today, users have widespread access to massively-parallel computing resources. For example, at LLNL, users have a choice of several platforms including ASCI White and an 11 TF Linux cluster. With ASCI Purple and BlueGene/L, LLNL will continue this trend into the foreseeable future. We have discovered that this unprecedented degree of parallelism exposes some new performance limitations of both applications and architectures. However, while investigating these limitations, we have found that analyzing and predicting performance at this scale is becoming increasingly difficult. First, performance analysis techniques must strike the appropriate balance between instrumentation resolution and overhead. Second, users and architects must gain insight from these potentially massive datasets. Third, to predict performance, designers need efficient performance modeling strategies for these large, complex applications. Our research addresses these challenges by focusing on scalable, innovative techniques for performance analysis and modeling. In this talk, I will present recent results that include the use of machine learning and multivariate statistical techniques to distill important information from massive performance datasets, a novel sampling technique for measuring MPI communication performance, and an application-driven performance modeling study of LLNL workloads.