Computer Science Seminar Series
Toward Automatic Parallelization of Spatial Computation for Computing Clusters
November 07, 3:00pm
Weir Hall, Room 235
Baoqiang Yan
In recent years, cluster computing resources have become freely available to a wide variety of scientific researchers in recent years. However, scientists are not necessarily skilled in writing efficient parallel code. The processing of spatial datasets is one area in which this problem is particularly acute. To address this issue, we are developing an API that helps the scientific user to easily write code that performs I/O efficiently and either performs efficient inter-compute node communication, or avoids it entirely.
In previous work we devised a method of distributing data amongst compute nodes that takes into account the manner in which the data is stored on disk, and also aggregates cluster I/O. This work was done in the context of ray casting, but the view direction was constrained to one of the major axes. In this paper, we extend our previous work to allow arbitrary rotation of the view direction around a major axis, meaning that spatial dependencies are present along two axes for a given view direction, making the problem much harder.