Speaker
Description
The conjugate gradient (CG) algorithm, employing mixed precision and even-odd preconditioning, is typically used when computing propagators for highly improved staggered quarks (HISQ). However, its performance degrades due to critical slowing down as the light quark mass approaches its physical value. In previous work, we addressed this issue using deflation, which significantly reduced solve times—achieving speedups of up to 10$\times$ over CG on the most challenging ensembles, albeit with a substantial setup cost to compute eigenvectors. In this work, we focus on a multigrid algorithm for HISQ propagators tuned to achieve optimal performance. Multigrid achieves even greater speedups—up to 17$\times$—nearly doubling the speedups achieved with deflation—while substantially reducing both the setup cost and the memory resources required. We present benchmarks using the MILC and QUDA software libraries on lattices up to $144^3×288$ (lattice spacing 0.04 fm) and quark masses ranging from the physical strange down to the physical light quark values. We compare CG, deflation, and multigrid, highlighting the trade-offs between setup cost, memory footprint, and overall solve-time savings.
| Parallel Session (for talks only) | Software development and machines |
|---|
