AstroGK Performance

AstroGK Strong Scaling

Here are the results of the same strong scaling test on a number of different leadership class supercomputing facilities around the country.

Despite its heavy reliance on nonlocal algorithms (implicit time-stepping, spectral schemes in x-space and v-space) AstroGK has proven to be highly scalable on current terascale computers. At fixed problem size, nearly perfect speedup is obtained from 1-1000 processors of the IBM SP RS/6000 at NERSC. Excellent strong scaling of the same problem is also found on other platforms, such as Seaborg, the IBM BlueGene/L at Argonne; Dawson, the NSF-funded Appleseed cluster at UCLA; Bassi, an IBM p575 POWER 5 system at NERSC; and Franklin, the 19,320 processor Cray XT4 at NERSC. Scaling to larger numbers of processors can be obtained by increasing the problem size.