The CCSD algorithm of Kobayashi and Rendell in the NWChem software package has been optimized to improve the performance and scalability of the code. This modification provided minimum 2-fold speedup in the CCSD iteration time and made the code scaling up to 20,000 XE6 nodes on NCSA Blue Waters, which is 89% of the CPU-only capacity of the system. We report a complete CCSD(T) calculation, running on 20,000 XE6 nodes and including all steps from initialization till termination, reaching 0.32 petaflop/sec performance level on the problem of relative stability of A- vs B-DNA single strand represented by GpC deoxydinucleotide monophosphate in 6-311++G** basis set. The section of perturbative triples reached the performance of 1.18 petaflop/sec. The optimization of CCSD algorithm employed the availability of 64 GB RAM on Blue Waters XE6 nodes to eliminate the communication bottleneck in accessing the atomic orbital-indexed doubly excited amplitudes. This work highlights the importance of availability of large on-node memory to the hardware architects of the future exaflop machines. Without adequate memory, coupled cluster codes may not be able to achieve high floating-point efficiency and scale to large node-counts.

Petascale optimization of the coupled cluster method in NWChem on NCSA Blue Waters

248th National Meeting (2014)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.