Scaling SciDAC Lattice QCD Software to GPU Enabled Parallel Systems
- Computing Techniques Seminars
- Data Movement Challenges in the Era of 100 Gbit and Terabit Networks
- Challenges and Opportunities in Large-Scale Storage Systems
- Scaling SciDAC Lattice QCD Software to GPU Enabled Parallel Systems
- Multi-core Optimizations for Synergia2
|Full Title:||Scaling SciDAC Lattice QCD Software to GPU Enabled Parallel Systems|
|Date & Time:||17 Apr 2012 at 14:00|
|Event Location:||Curia II|
|Event Info:||Speaker: Frank Winter, Institute for Particle Physics Theory, The University of Edinburgh.
Lattice QCD, the tool to explore the non-perturbative regime of QCD, is amongst the most computationally demanding disciplines in theoretical physics. SciDAC's QDP++, a C++ library for lattice quantum field theories, has proven to scale Lattice QCD applications to CPU based capability machines. Current trends in the high-performance computing landscape show an increasing dominance of extreme-scale hybrid-computing architectures enabled by GPU accelerators. This work introduces QDP-JIT which supports computations on NVIDIA GPU enabled parallel systems while maintaining the SciDAC QCD data parallel application interface. Single expression evaluations are off-loaded to the accelerators using Just-In-Time (JIT) compilation techniques and employing a caching mechanism automates the memory management.
Frank Winter is a Post-doctoral researcher in Lattice QCD software development and optimization at the Institute for Particle Physics Theory, The University of Edinburgh. He received his Ph.D. in Theoretical Physics 2011 from the University of Regensburg, Germany. He has worked at the John von Neumann Institute for Computing, DESY-Zeuthen developing and deploying lattice software for the QPACE parallel supercomputer based on the Cell processor.