| 摘要 |
A data structure, the distribution, may be provided to track the desired and/or actual location of computations and data that range over a multidimensional rectangular index space in a parallel computing system. Examples of such iteration spaces include multidimensional arrays and counted loop nests. These distribution data structures may be used in conjunction with locality-guided work stealing and may provide a structured way to track load balancing decisions so they can be reproduced in related computations, thus maintaining locality of reference. They may allow computations to be tied to array layout, and may allow iteration over subspaces of an index space in a manner consistent with the layout of the space itself. Distributions may provide a mechanism to describe computations in a manner that is oblivious to precise machine size or structure. Programming language constructs and/or library functions may support the implementation and use of these distribution data structures.
|
