GPU optimizations for a production molecular docking code
Landaverde, Raphael J.
MetadataShow full item record
Scientists have always felt the desire to perform computationally intensive tasks that surpass the capabilities of conventional single core computers. As a result of this trend, Graphics Processing Units (GPUs) have come to be increasingly used for general computation in scientific research. This field of GPU acceleration is now a vast and mature discipline. Molecular docking, the modeling of the interactions between two molecules, is a particularly computationally intensive task that has been the subject of research for many years. It is a critical simulation tool used for the screening of protein compounds for drug design and in research of the nature of life itself. The PIPER molecular docking program was previously accelerated using GPUs, achieving a notable speedup over conventional single core implementation. Since its original release the development of the CPU based PIPER has not ceased, and it is now a mature and fast parallel code. The GPU version, however, still contains many potential points for optimization. In the current work, we present a new version of GPU PIPER that attains a 3.3x speedup over a parallel MPI version of PIPER running on an 8 core machine and using the optimized Intel Math Kernel Library. We achieve this speedup by optimizing existing kernels for modern GPU architectures and migrating critical code segments to the GPU. In particular, we both improve the runtime of the filtering and scoring stages by more than an order of magnitude, and move all molecular data permanently to the GPU to improve data locality. This new speedup is obtained while retaining a computational accuracy virtually identical to the CPU based version. We also demonstrate that, due to the algorithmic dependencies of the PIPER algorithm on the 3D Fast Fourier Transform, our GPU PIPER will likely remain proportionally faster than equivalent CPU based implementations, and with little room for further optimizations. This new GPU accelerated version of PIPER is integrated as part of the ClusPro molecular docking and analysis server at Boston University. ClusPro has over 4000 registered users and more than 50000 jobs run over the past 4 years.
Thesis (M.Sc.Eng.) -- Boston University
RightsThis thesis is being made available in OpenBU by permission of its author, and is available for research purposes only. All rights are reserved to the author.