PNNL

Scientists Jeff Daily, Abhinav Vishnu, and Bruce Palmer, all from the ACMD Division High Performance Computing group at PNNL, served as the core team for a new release of the Global Arrays (GA) toolkit, known as Version 5.5. GA 5.5 provides additional support and bug fixes for the parallel Partitioned Global Address Space (PGAS) programing model.

(L to R): Daily, Vishnu, and Palmer worked closely with Intel on the recent update to Global Arrays. Enlarge Image.

GA 5.5 incorporates support for libfabric (https://ofiwg.github.io/libfabric/), which helps meet performance and scalability requirements of high-performance applications, such as PGAS programming models (like GA), Message Passing Interface (MPI) libraries, and enterprise applications running in tightly coupled network environments. The updates to GA 5.5 resulted from a coordinated effort between the GA team and Intel Corp. Along with incorporating support for libfabric, the update added native support for the Intel Omni-Path high-performance communication architecture and applied numerous bug fixes since the previous GA 5.4 release to both Version 5.5 and the ga-5-4 release branch of GA’s subversion repository.

Originally developed in the late 1990s at PNNL, the GA toolkit offers diverse libraries employed within many applications, including quantum chemistry and molecular dynamics codes (notably, NWChem), as well as those used for computational fluid dynamics, atmospheric sciences, astrophysics, and bioinformatics.

“This was a significant effort from Intel to work with us on the libfabric, and eventual Intel Omni-Path, support,” Daily explained. “Had we not refactored our Global Arrays one-sided communication library, ComEx, a few years ago to make it easier to port to new systems, this would not have been possible. Now that our code is much easier to integrate with, we envision more collaborations like this in the future.”

Download information for GA 5.5. and the GA subversion repository is available here.