Computer Science
Reducing the performance gap
The SciDAC Performance Evaluation Research Center (PERC) is working toward reducing the gap between peak and sustained performance on high-end computers by developing hardware and software tools to monitor, model, and control performance. In the past year, PERC has analyzed and optimized application codes in astrophysics, lattice gauge theory, climate modeling, applied mathematics, and other disciplines, significantly improving the performance and scalability of the codes. For example, improved compiler annotation technology (Active Harmony) achieved a 15% reduction in run time on the POP ocean modeling code.
D. H. Bailey, B. de Supinski, J. Dongarra, et al., "Performance technologies for peta-scale systems," http://perc.nersc.gov/docs/PERC-white-aes15.doc. ASCR-MICS, SciDAC
Integrating the Globus Toolkit with firewalls
The Globus Toolkit is a collection of clients and services that enable Grid computing. Each component of the toolkit has its own network traffic characteristics that must be considered when deploying the toolkit at a site with a firewall. Members of the Globus Project and the DOE Science Grid engineering team have developed requirements and guidance for firewall administrators, describing the network traffic generated by the Globus Toolkit services, how the Toolkit’s use of the ephemeral port range can be controlled, and firewall requirements for client and server sites. NERSC provided a firewall testbed that was used for verifying much of the information in this document.
V. Welch et al., “Globus Toolkit Firewall Requirements Version 5,” July 22, 2003.
Preconditioning inexact Newton algorithms
Inexact Newton algorithms are commonly used for solving large sparse nonlinear system of equations F(u*) = 0. Even with global strategies such as line search or trust region, the methods often stagnate at local minima of ||F||, especially for problems with unbalanced nonlinearities, because the methods do not have built-in machinery to deal with the unbalanced nonlinearities. Cai and Keyes studied a nonlinear additive Schwarz-based parallel nonlinear preconditioner; they showed numerically that the new method converges well even for some difficult problems, such as high Reynolds number flows, where a traditional inexact Newton method fails.
X.-C. Cai and D. E. Keyes, "Nonlinearly preconditioned inexact Newton algorithms," SIAM J. Sci. Comp. 24, 183 (2002). ASCR-MICS, SciDAC, NSF, NASA
Developing an alternative programming model
While MPI is currently the de facto standard for programming supercomputers, several federal agencies are investing in alternative models to simplify programming. Yelick and collaborators are investigating UPC, a “partitioned global address space” language. They developed and released an open-source version of the full UPC compiler, which translates UPC to C with calls to a newly developed UPC runtime layer. The open-source compiler has low overhead for basic operations on shared data; it is competitive with, and sometimes faster than, the commercial compiler.
W. Chen, D. Bonachea, J. Duell, P. Husbands, C. Iancu, and K. Yelick, “A performance analysis of the Berkeley UPC compiler,” 17th Annual International Conference on Supercomputing (ICS), 2003. ASCR-MICS, DOD, NSF