USQCD Software Report Steven Gottlieb Indiana University (MILC - - PowerPoint PPT Presentation

Steven Gottlieb Indiana University (MILC & Fermilab Lattice/MILC Collaborations)

Fermilab Batavia, IL October 18, 2013

USQCD Software Report

• S. Gottlieb, Fermilab, 10-18-13

Funding/Personnel

✦ Arizona, Indiana, NCSA, Utah applied to both NSF

Physics at the Information Frontier (PIF) program and SciDAC for funding of postdocs to work on software.

✦ DOE and NSF decided they did not like splitting

support so, Arizona & Utah are funded by DOE, and NCSA & Indiana are funded by NSF.

✦ Work under PIF was delayed because Guochun Shi left

NCSA for Google (just before our funding was assured), and Indiana was not able to hire a postdoc in the first Fall.

✦ NCSA: Nuno Cardoso (Craig Steffen) ✦ Indiana: Mathias Wagner

2

• S. Gottlieb, Fermilab, 10-18-13

Performance Modeling

✦ Long history of benchmarking MILC code, mostly in

weak scaling mode

✦ For Blue Waters, there was a strong push for

performance modeling prior to delivery. I worked on that while on sabbatical at NCSA with Greg Bauer and Torsten Hoefler

✦ Holy grail of performance modeling:

• describe hardware (what level?)
• predict performance of single and multinode code

✦ Semi-empirical approach

• measure single node performance
• measure message passing performance

3

• S. Gottlieb, Fermilab, 10-18-13

Performance Modeling...

✦ Performance Modeling and Comparative Analysis of

the MILC Lattice QCD Application su3_rmd, CCGrid 2012

• prepared in anticipation of IBM Blue Waters
• actual Blue Waters is a Cray

✦ Have encouraged Nuno to get involved with this for

GPU computing on Blue Waters

✦ He has started with inverter benchmarking

• needs more of my attention

4

• S. Gottlieb, Fermilab, 10-18-13

SU(3) X U(1)

✦ We have been interested in developing a code with

dynamical QED for some time.

✦ Ran Zhou has been working on a reference CPU

implementation.

• Ran now at Fermilab.

✦ Expect Mathias Wagner to work on GPU

implementation.

• He has a lot of experience with RHMC code from Bielefeld GPU

cluster development.

5