Statistics – Computation
Scientific paper
Dec 2010
adsabs.harvard.edu/cgi-bin/nph-data_query?bibcode=2010agufmin41a1357k&link_type=abstract
American Geophysical Union, Fall Meeting 2010, abstract #IN41A-1357
Statistics
Computation
[0545] Computational Geophysics / Modeling, [1906] Informatics / Computational Models, Algorithms, [1932] Informatics / High-Performance Computing, [5460] Planetary Sciences: Solid Surface Planets / Physical Properties Of Materials
Scientific paper
Discrete element method (DEM) simulation is used for the modeling of mobility and excavation on Mars and the Moon, due to many prohibitive limitations of continuum models. One of the most computationally expensive parts of all DEM codes is contact detection between different particles and objects in the system. Taking a naive approach, the operation costs grow as a square of the number of elements. This is prohibitive even for a small number of particles. Instead, most of the impossible contacts are filtered by broad phase contact detection systems. However, such systems are generally hard to run on parallel heterogeneous systems due to many conditionals and to large global memory transactions. In this work, we report on a new broad phase contact detection system, suitable for heterogeneous computational systems such as the IBM Cell and CPU/GPU systems. The approach is based on space partitioning and sweep and prune methods. The particle space is divided by regular columns, where every column is filled with particles. Each particle belongs to up to 4 columns. A sweep and prune method is used in every column. Due to every column containing number of particles proportional to a cubic root of total number of particles, the column fits in small messages or in small memory chunks and can be independently processed on a small operational unit, such as an SPU on the IBM Cell processor. The special method used to enforce the contact list produced by this procedure, so that it contain no duplicates of collisions. Further parallelization is allowed on other stages due to this feature of the method, as well. We discuss the usage and performance of this method on an IBM Cell QS22 cluster with 12 blades, available at Arctic Region Supercomputing Center. We compare to CPU based supercomputers. To make the code portable between different systems, we abstract the parallel level to a separate software library. The main code is "aware" it needs to run in a parallel mode. However, the parallel model is chosen by the library. The whole code is written mostly in C using literate programming approach (NOWEB) with some parts written in Lua. It uses HDF for I/O.
Johnson Jeffrey B.
Kulchitsky A. V.
No associations
LandOfFree
Hybrid Broad Phase Contact Detection Method for Lunar/Mars Regolith Modeling Designed for Use on Heterogeneous Computer Systems does not yet have a rating. At this time, there are no reviews or comments for this scientific paper.
If you have personal experience with Hybrid Broad Phase Contact Detection Method for Lunar/Mars Regolith Modeling Designed for Use on Heterogeneous Computer Systems, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Hybrid Broad Phase Contact Detection Method for Lunar/Mars Regolith Modeling Designed for Use on Heterogeneous Computer Systems will most certainly appreciate the feedback.
Profile ID: LFWR-SCP-O-1502295