First LQCD Physics Runs with MILC and P4RHMC

Abstract

An initial series of physics LQCD runs were submitted to the BG/L science bank with the milc and p4rhmc. Both runs were for lattice dimensions of 32{sup 2} x 8. The p4 calculation was performed with v2.0 QMP{_}MPI.X (semioptomized p4 code using qmp over mpi) and milc v7.2, also using RHMC, but not specifically optimized for BlueGene. Calculations were performed along lines of constant physics, with the light quark masses 2-3 times their physics values and the strange quark mass set by m{sub ud} = 0.1m{sub s}. Job submissions was performed using the standard milc and p4 scripts provided on the ubgl cluster. Initial thermalized lattices for each code were also provided in this way. The only modifications for running on BG/L were to the directory names and the mT parameter which determines job durations (24 hrs on BG/L vs. 4 hrs on ubgl). The milc scripts were set to resubmit themselves 10 times, and the p4 scripts were submitted serially using the ''psub -d'' job dependency option. The runp4rhmc.tcsh could not be used to resubmit due to the 30m time limit imposed on interactive jobs. Most jobs were submitted to the smallest, 512 node partitions, but both codes could also run on the 1024 node partitions with a gain of only 30-50%. The majority of jobs ran without error. Stalled jobs were often indicative of a communication gap within a partition that LC was able to fix quickly. On some occasion a zero-length lattice file was deleted to allow jobs to restart successfully. Approximately 1000 trajectories were calculated for each beta value, see Table . The analysis was performed with the standard analysis scripts for each code, make{_}summary.pl for milc and analysis.tcsh for p4rhmc. All lattices, log files, and job submission scripts have been archived to permanent storage for subsequent analysis.

Abstract

Related Papers