Multi-mass solvers for lattice QCD on GPUs

TL;DR

This paper demonstrates an efficient memory-lean approach using single-mass inverters for multi-mass lattice QCD calculations on GPUs, outperforming traditional multi-shift algorithms in speed and memory usage.

Contribution

It introduces a novel method employing single-mass inverters for multi-mass problems, reducing memory requirements and doubling performance compared to multi-shift algorithms.

Findings

Single-mass inverter requires less memory.

Single-mass inverter outperforms multi-shift by a factor of two.

Method is effective for Wilson fermions on GPUs.

Abstract

Graphical Processing Units (GPUs) are more and more frequently used for lattice QCD calculations. Lattice studies often require computing the quark propagators for several masses. These systems can be solved using multi-shift inverters but these algorithms are memory intensive which limits the size of the problem that can be solved using GPUs. In this paper, we show how to efficiently use a memory-lean single-mass inverter to solve multi-mass problems. We focus on the BiCGstab algorithm for Wilson fermions and show that the single-mass inverter not only requires less memory but also outperforms the multi-shift variant by a factor of two.