The Numerical Algorithms Group has announced new HPC performance milestones.
These include up to four-times better performance with multicore optimisation for materials science and quantum Monte Carlo applications, and reductions of up to 25 per cent in runtimes with I/O tuning for an ocean-modelling application.
These are the early results of NAG's Distributed Computational Science and Engineering (DCSE) support program for Hector (the UK's national supercomputing facility), which now consists of more than 30 dedicated application-optimisation projects complementing the traditional HPC user-support provided by NAG.
NAG enhanced Castep - a key materials science code used by academic researchers and industry - with band-parallelism to allow the code to scale to more than 1,000 cores.
The speed of Castep on a fixed number of cores was also improved by up to four times on the original, representing a potential saving of more than USD3m of computing resources over the remainder of the Hector service.
The Castep project showed the collaborative nature of the DCSE program, with the University of York undertaking the core development in conjunction with NAG HPC staff and the Science and Technology Facilities Council.
In another project, an ocean-modelling application known as Nemo (Nucleus for European Modelling of the Ocean) underwent optimisation, including I/O techniques and variable resolution approaches to run 25 per cent faster on relevant-use cases.
This represented a USD600,000 saving in computing resources for that project, with potentially multi-million dollar savings across all Nemo users.
The project was performed by a collaboration of the National Oceanography Centre and the University of Edinburgh working with NAG HPC staff.
A third project optimised a quantum Monte Carlo code (Casino) for better performance on multicore nodes by introducing shared memory techniques and hierarchical parallelism.
This resulted in performance gains of up to four-times on quad-core nodes and further performance gains from I/O optimisations for simulations using more than 10,000 cores.
Following NAG's work, the scientists were able to run on 40,000 cores of the Jaguar Petaflops supercomputer at Oak Ridge National Laboratory.
This DCSE project was undertaken by NAG HPC staff working with users at University College London and is estimated to have saved the researchers around USD1m in computing resources on Hector.
NAG's DCSE projects enhance the performance of dozens of community HPC applications used on Hector and all performance improvements are fed back into the community so that non-Hector users can benefit too.
Further examples can be found on the NAG website.
