Statistical analysis of simulation-generated time series: Systolic vs. semi-systolic correlation on the Connection Machine

Abstract

Autocorrelation becomes an increasingly important tool to verify improvements in the state of the simulational art in Latice Gauge Theory. Semi-systolic and full-systolic algorithms are presented which are intensively used for correlation computations on the Connection Machine CM-2. The semi-systolic algorithm makes use of an intrinsic, microprogrammed global-add reduction function which is implemented extremely well on the Connection Machine. Nevertheless, the full-systolic correlation algorithm which makes use only of local communication and computation operations turns out to be substantially superior to the semi-systolic scheme whose basic step involved a non-local sum computation that extends over the entire machine.