Cobalt Solutions, LLC set out to answer these questions by performing recent scalability testing of Cobalt which demonstrated superlinear scalability into the tens of thousands of processors.
The previous scaling test of Cobalt was in 2005 and utilized 4,000 processors - the maximum number available at the time on DoD systems. The recent testing utilized up to 21,856 processors – the maximum number available defined by a queue system - on a Cray XE6 located at Wright-Patterson AFB.
The results, contained in the paper entitled Can Legacy Codes Scale on Tens of Thousands of PEs or Do We Need to Reinvent the Wheel?, demonstrate that Cobalt produces superlinear speed-up on a 49 million cell grid run on a range of processors from 96 up to 21,856. This comes out to 2,200 cells-per-PE at the highest number of PEs. A 3 million cell grid and a 10 million cell grid were also run and exhibited similar scaling.
The current versions of Cobalt, V5.2 and B5.11, both available on most DoD HPC machines, have proven to be highly scalable, scaling well on tens of thousands of processors. It has been shown that Cobalt clearly has the potential to scale well on hundreds of thousands of processors, given a grid size of roughly 85 million cells or larger. Additionally, it has been shown that scalability efficiency for Cobalt can be roughly predicted based on a cells-per-PE basis.