Ready on day 1

To ensure that scientists and engineers can achieve sustained petaflop performance on Blue Waters from day one, intensive work is under way now to port, optimize, and scale a range of applications to effectively use the system’s more than 200,000 processors. Scientists and engineers who use these applications are working closely with computing experts, including staff from IBM, through Petascale Application Collaboration Teams (PACTs) to achieve this unprecedented level of performance.

To ensure that scientists and engineers can achieve sustained petaflop performance on Blue Waters from day one, intensive work is under way now to port, optimize, and scale a range of applications to effectively use the system’s more than 200,000 processors. Scientists and engineers who use these applications are working closely with computing experts, including staff from IBM, through Petascale Application Collaboration Teams (PACTs) to achieve this unprecedented level of performance.

The current PACTs focus on challenges in lattice quantum chromodynamics, biomolecular dynamics, and turbulent gas dynamics that were outlined by the National Science Foundation (NSF) in the solicitation that led to the Blue Waters project. In each discipline, NSF has described a large-scale problem and specified a time to solution that represents a tremendous leap beyond today’s capabilities. While we are both members of the turbulence PACT, we are also closely involved in other Blue Waters application development efforts.

Achieving the performance milestones is an intrinsic and vital part of the Blue Waters project. Our goal is not simply to build and deploy a powerful machine. It is essential for the machine and the targeted applications to work together to enable discovery and innovation.

Scaling to large numbers of processors requires techniques and approaches that are specific to each application. For example, it may be necessary to modify or change algorithms, use finer grained decomposition, eliminate replication of data structures and long critical paths, employ dynamic load balancing, etc.

To prepare the model problems for Blue Waters, they are being benchmarked on large processor counts on currently available systems with petascale peak speeds, such as the BlueGene/P at Argonne National Laboratory and the Cray XT5 at Oak Ridge National Laboratory. We are studying potential bottlenecks to scalability as they emerge, with code and algorithm modifications being implemented as needed.

In addition, the performance of modules in these codes on the level of a single Blue Waters compute node is being predicted using Mambo, IBM’s application performance simulator for the POWER7 processor. Using these single-node predictions, the overall performance of applications running on the entire system, including the internodal communication network, is being estimated using BigSim, which was developed by a University of Illinois team led by computer science professor Laxmikant Kale, or using the analytical network modeling approach developed by Adolfy Hoisie and Darren Kerbyson at Los Alamos National Laboratory. In this way, we can make modifications that are needed to enable the applications to use the full capability that Blue Waters will provide when it becomes available in 2011.

While each PACT functions independently, members of the different teams also meet weekly in conjunction with performance experts from IBM to discuss issues and potential solutions. Lessons learned in optimizing one application will be applied to others as appropriate.

A solicitation is being prepared for one or two new PACTs to be formed soon, with more to follow each year, to prepare additional applications for Blue Waters. The codes on which these PACTs focus will:

• have a high potential for petascale-enabled scientific or engineering breakthroughs,
• already scale to thousands of cores,
• have the potential to effectively use the Blue Waters architecture,
• represent a broad range of science and engineering application types and algorithms,
• span a wide range of available programming models,
• have or potentially have analytic internode performance models.

If you are working on a code that fits these criteria, we welcome you to contact the Blue Waters Consulting Office: help+bw@ncsa.illinois.edu.

Disclaimer: Due to changes in website systems, we've adjusted archived content to fit the present-day site and the articles will not appear in their original published format. Formatting, header information, photographs and other illustrations are not available in archived articles.