Engineering Mechanics Institute Conference 2013

Full Program »

Mapping the PolyTop framework to a multi-node GPU cluster

DOC Abstract
View File
doc
97KB

Leonardo S. Duarte
PUC-Rio
Brazil

Waldemar Celes
PUC-Rio
Brazil

Ivan F.M. Menezes
PUC-Rio
Brazil

Glaucio H. Paulino
Illinois University
United States

Abstract:
The PolyTop framework was improved to be executed successfully on a multi-node GPU cluster by using CUDA programming model. PolyTop is an efficient Matlab code for topology optimization using unstructured polygonal finite element meshes. Good results have already been achieved with the PolyTop on a single GPU. However, the high amount of memory storage required for the mapping matrix and the computational power necessary to the solution of the linear system of equations bring the use of many GPUs indispensable for large-scale simulations. For multi-GPU computation, decreasing data communication time is the key idea to enhance parallel efficiency. Domain partitioning method is used to distribute the computational load of an element-by-element iterative solver, since this is the main bottleneck of the PolyTop performance. Comparisons and analyses were made among data layouts and overlapping techniques between computation and communication, in order to hide the data transfer time between GPU nodes. Asynchronous streams enable simultaneously computation during the communication process. Examples are provided to demonstrate the speedups achieved and capability of the code to simulate large problems.

 

Powered by OpenConf®
Copyright ©2002-2012 Zakon Group LLC