1 of 1

Graph Methods for Lattice QCD CalculationsWith the Nuclear Physics (NP) LQCD SciDAC5 Partnership program

1

Scientific Achievement

We propose and develop novel algorithms for scheduling and distributing computation of the correlation functions on the accelerators. Our methods drastically reduce the memory footprint of the tensor contraction operations and data transfers across accelerators, opening up the path for more complex lattice QCD simulations containing several hadrons.

Significance and Impact

Computational challenges in calculating correlation functions in lattice QCD simulations limit the size of the hadron systems that can be simulated. Solving these challenges with novel combinatorial graph models, we significantly reduce the time-to-solution, improve scalability, and increase the scale of the systems that can be simulated within the Redstar framework.

The number of device to device operations of the Redstar’s scheduler (micco) and the developed graph-partitioning model (sbgp) for four different hadron systems on two and four accelerators. Reducing data movement among multiple accelerators is key to scalability and faster simulations.

(numbers on the last bar on each subfigure is the reduction factor)

Technical Approach

  • We devised an efficient algorithm for scheduling contractions onto an accelerator that aims to minimize memory footprint in order to avoid evictions.
  • We developed a novel graph-partitioning model to distribute contractions among multiple accelerators to tackle the data movement costs and hence improve Redstar’s scalability.

PI(s)/Facility Lead(s): Aydin Buluc, Oguz Selvitopi

Collaborating Institutions: Jefferson Lab, Lawrence Berkeley National Laboratory, University of Utah

ASCR Program: SciDAC NP

ASCR PM: Lali Chatterjee, Kalyan Perumalla

Code Developed or Datasets: Redstar

2.6x

16.2x

368x

21.9x