Show simple item record

dc.contributor.authorKhan, MD. Ashfaquzzamanen_US
dc.date.accessioned2015-08-07T03:17:22Z
dc.date.available2015-08-07T03:17:22Z
dc.date.issued2013en_US
dc.date.submitted2013en_US
dc.identifier.other(ALMA)contempen_US
dc.identifier.urihttps://hdl.handle.net/2144/12792
dc.descriptionThesis (Ph.D.)--Boston University PLEASE NOTE: Boston University Libraries did not receive an Authorization To Manage form for this thesis or dissertation. It is therefore not openly accessible, though it may be available by request. If you are the author or principal advisor of this work and would like to request open access for it, please contact us at open-help@bu.edu. Thank you.en_US
dc.description.abstractWhile Molecular Dynamics Simulation (MD) uses a large fraction of the world's High Performance Compute cycles, the modeling of many physical phenomena remains far out of reach. Improving the cost-effectiveness of MD has therefore received much attention, especially in using accelerators or modifying the computation itself. While both approaches have demonstrated great potential, scalability has emerged as a critical common challenge. The goal of this research is to study this issue and develop MD solutions that not only achieve substantial acceleration but also remain scalable. In the first part of this research, we focus on Discrete Molecular Dynamics Simulation (DMD)., which achieves high performance by simplifying the underlying computation by converting it into a Discrete Event Simulation (DES). In addition to the inherent serial nature of DES, causality issues make DMD a notorious target for parallelization. We propose a parallel version of DMD that, unlike any previous work, uses task decomposition and efficient synchronization and achieves more than 8.5x speed-up for 3D physical systems on a 12 core processor, with potential for further strong scaling. The second part of this research focuses on FPGA acceleration of timestep-driven MD. We first enhance an existing FPGA kernel to take advantage of the Block RAM architecture of FPGAs. This results in a 50% improvement in speed-up, without sacrificing simulation quality. We then parallelize the design targeting multiple on-board FPGA cores. We combine this with software pipelining and careful load distribution at the application level to achieve a 3.37x speedup over its CPU counterpart. In the third part we create a framework that integrates the FPGA accelerator into a prominent MD package called NAMD. This framework allows users to switch between the actual accelerator and a simulated version, and provides a means to study different characteristics, such as the communication pattern, of such an accelerated system. Using this framework, we identify the drawbacks of the current FPGA kernel and provide guidelines for future designs. In addition, the integrated design achieves 2.22x speed-up over a quad-core CPU, making it the first ever FPCA-accelerated full-parallel MD package to achieve a positive end-to-end speed-up.en_US
dc.language.isoen_USen_US
dc.publisherBoston Universityen_US
dc.titleScalable molecular dynamics simulation using FPGAs and multicore processorsen_US
dc.typeThesis/Dissertationen_US
etd.degree.nameDoctor of Philosophyen_US
etd.degree.leveldoctoralen_US
etd.degree.disciplineComputer Engineeringen_US
etd.degree.grantorBoston Universityen_US


This item appears in the following Collection(s)

Show simple item record