In 2013, the Argonne Leadership Computing Facility (ALCF) will be home to a new production resource - Mira, a 10-petaflops Blue Gene/Q (BG/Q) supercomputer - currently ranked #3 in the Top500. The BG/Q is the 3rd generation of the Blue Gene platform and continues the Blue Gene design philosophy, providing a among other things a leading FLOP per Watt ratio. The BG/Q however supports a number of novel features that improve the per-node and network performance in comparison to its predecessors. The BG/Q platform also allows for significantly more cores than previous BG/P systems - Mira contains over 768k cores which is six times more than Intrepid.

Experience has shown that applications running on the BG/P can readily transition to the BG/Q, and with moderate tuning efforts can obtain a significant fraction of the 15x per node peak FLOP rate improvement from BG/P to BG/Q. The ALCF Performance Engineering Team participated in two separate, but equally important projects during the design and development of the BG/Q. In the first project, the team tracked the performance of 10 key ALCF applications, starting with extracted kernels on cycle-accurate simulators and FPGAs to full applications on Early Access hardware. In the second project, the team worked with IBM on the requirements and design of the BG/Q hardware performance counter API to incorporate lessons learned from the BG/P and other HPC architectures. This talk will provide details on the two projects, focusing on the performance analysis of the 10 applications on Mira. The performance counter data from these applications will offer valuable insight on the right choice of programming models to improve single-node performance on the BG/Q.