NGS: Collaborative Research: Performance Measurement & Modeling of Deep Hierarchy Systems

University Of California-San Diego, La Jolla CA

Investigators

Abstract

CNS-0406312 The goal of this project is to advance methods for understanding the performance of scientific applications. Because of current architectural trends the performance of applications depends largely on their interactions with memory and interconnect subsystems of today's large-scale HEC (High-End Computing) platforms. The dynamics of these interactions are increasingly complex and subtle; therefore means for understanding them are needed so that applications can be better tuned to improve time-to-solution, and future machines can be designed to better meet the needs of applications. The Principal Investigators (PIs) will carry out a program of research and development to provide the computational science community with EMPS (Environment for Memory Performance Studies), an integrated simulation environment for posing and answering "what if" performance questions. Capabilities that this project will provide include: 1. Tools for capturing program's performance data with special emphasis on memory and communications. 2. Tools for abstracting and summarizing the essential performance behaviors of applications from the performance data. 3. Facilities for forming and evaluating performance models to identify the performance sensitivity of applications to changes in machine, or algorithm, or implementation. 4. An EMPS API to enable plug-and play between performance tool components the PIs develop and also allowing the environment to be extensible to include components developed by others. Intellectual Merit: As part of this work the PIs will develop a science for understanding the performance of applications running on deep and wide memory hierarchy machines. They will answer the question "what are the factors that affect the performance of scientific applications on such machines?" Going a step further, they will quantify the effects of performance factors, thus increasing the exactitude of performance models for these architectures. They will also develop predictive methods to evaluate how the performance of applications would benefit from new architectural features, and how much performance would improve if different algorithms and/or optimizations were used- thus to evaluate options and tradeoffs prior to carrying out expensive development efforts. Broader Impact: The proposed work will develop a variety of techniques that will permit application scientists to better understand their applications and allow computer scientists to develop computers that are better able to accommodate applications. This will potentially allow faster and better scientific simulations that enable new research to be done in many scientific disciplines. In addition by providing a close working relationship between a university, an NSF center, and a major HPC vendor this project will produce students who are trained in the many facets that impact HPC system development.