The BlueGene/L Supercomputer and Quantum ChromoDynamics

In summary our update contains: (1) Perfect speedup sustaining 19.3% of peak for the Wilson D D-slash Dirac operator. (2) Measurements of the full Conjugate Gradient (CG) inverter that inverts the Dirac operator. The CG inverter contains two global sums over the entire machine. Nevertheless, our measurements retain perfect speedup scaling demonstrating the robustness of our methods. (3) We ran on the largest BG/L system, the LLNL 64 rack BG/L supercomputer, and obtained a sustained speed of 59.1 TFlops. Furthermore, the speedup scaling of the Dirac operator and of the CG inverter are perfect all the way up to the full size of the machine, 131,072 cores (please see Figure II). The local lattice is rather small (4 x 4 x 4 x 16) while the total lattice has been a lattice QCD vision for thermodynamic studies (a total of 128 x 128 x 256 x 32 lattice sites). This speed is about five times larger compared to the speed we quoted in our submission. As we have pointed out in our paper QCD is notoriously sensitive to network and memory latencies, has a relatively high communication to computation ratio which can not be overlapped in BGL in virtual node mode, and as an application is in a class of its own. The above results are thrilling to us and a 30 year long dream for lattice QCD