Clover Action for Blue Gene-Q and Iterative solvers for DWF

TL;DR

This paper optimizes Clover fermion action for Blue Gene-Q and compares iterative solvers for Domain Wall Fermion, introducing a faster multi-shift MCR algorithm for lattice QCD simulations.

Contribution

It presents the first optimization of Clover action on Blue Gene-Q and introduces a new multi-shift MCR solver that outperforms existing methods.

Findings

Clover action achieved up to 29.1% efficiency in single precision.

Modified Conjugate Residual (MCR) is the most efficient solver for DWF.

Multi-shift MCR is 18.5% faster than multi-shift CG for rational function evaluation.

Abstract

In Lattice QCD, a major challenge in simulating physical quarks is the computational complexity of these simulations. In this proceeding, we describe the optimisation of Clover fermion action for Blue gene-Q architecture and how different iterative solvers behave for Domain Wall Fermion action. We find that the optimised Clover term achieved a maximum efficiency of 29.1% and 20.2% for single and double precision respectively for iterative Conjugate Gradient solver. For Domain Wall Fermion action (DWF) we found that Modified Conjugate Residual(MCR) as the most efficient solver compared to CG and GCR. We have developed a new multi-shift MCR algorithm that is 18.5% faster compared to multi-shift CG for the evaluation of rational functions in RHMC.