Evaluation Optimization and Application of Execution Models for Exascale Computing
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
IEEE Computer
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Physical Chemistry Chemical Physics
Abstract not provided.
Abstract not provided.
Abstract not provided.
IEEE SIGMETRICS PER
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Journal of Computational Chemistry
Abstract not provided.
Abstract not provided.
Journal of Computational Chemistry
Sharing low-level functionality between software packages enables more rapid development of new capabilities and reduces the duplication of work among development groups. Using the component approach advocated by the Common Component Architecture Forum, we have designed a flexible interface for sharing integrals between quantum chemistry codes. Implementation of these interfaces has been undertaken within the Massively Parallel Quantum Chemistry package, exposing both the IntV3 and Cints/Libint integrals packages to component applications. Benchmark timings for Hartree-Fock calculations demonstrate that the overhead due to the added interface code varies significantly, from less than 1% for small molecules with large basis sets to nearly 10% for larger molecules with smaller basis sets. Correlated calculations and density functional approaches encounter less severe performance overheads of less than 5%. While these overheads are acceptable, additional performance losses occur when arbitrary implementation details, such as integral ordering within buffers, must be handled. Integral reordering is observed to add an additional overhead as large as 12%; hence, a common standard for such implementation details is desired for optimal performance. © 2007 Wiley Periodicals, Inc.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
By means of coupled-cluster theory, molecular properties can be computed with an accuracy often exceeding that of experiment. The high-degree polynomial scaling of the coupled-cluster method, however, remains a major obstacle in the accurate theoretical treatment of mainstream chemical problems, despite tremendous progress in computer architectures. Although it has long been recognized that this super-linear scaling is non-physical, the development of efficient reduced-scaling algorithms for massively parallel computers has not been realized. We here present a locally correlated, reduced-scaling, massively parallel coupled-cluster algorithm. A sparse data representation for handling distributed, sparse multidimensional arrays has been implemented along with a set of generalized contraction routines capable of handling such arrays. The parallel implementation entails a coarse-grained parallelization, reducing interprocessor communication and distributing the largest data arrays but replicating as many arrays as possible without introducing memory bottlenecks. The performance of the algorithm is illustrated by several series of runs for glycine chains using a Linux cluster with an InfiniBand interconnect.
Proposed for publication in ClusterWorld.
Abstract not provided.