by M. Kistler,
J. Gunnels,
D. Brokenshire,
and B. Benton
We describe the challenges and opportunities we encountered when
developing a hybrid version of the Linpack benchmark for the Los Alamos National Laboratory Roadrunner supercomputing system,
which combines traditional x86-64 host processors with IBM
PowerXCell™ 8i accelerator processors. The challenges included
determining the proper division of the host and accelerator roles in
the computation, transfer of data between the host and accelerator
memory domains, alignment of data for communication and
computation, and data format differences between the two
processors. We also describe our approach to modeling the
performance of the hybrid system and compare our performance
estimates to witnessed performance on the system at different
scales and levels of memory consumption. Through careful
attention to these issues, we have produced a hybrid version of the
Linpack benchmark for the Roadrunner system that achieves
77.8% of peak performance on a single compute node and 74.6% of
peak performance over the entire system, making this system the
first to achieve a Linpack result exceeding one petaflops (1015
floating-point operations per second).
PDF
