Poor performance for large element selections

Hdf-forum archives
1

I'm trying to use H5S_select_elements to implement a parallel code that only
reads in the elements that are in the current partition. This works fine
when the number of elements in a partition is small but due to the linked
list implementation of the point selection, destruction of the dataspace
takes forever when I have on the order of 1 million+ point selections. Does
anyone have any ideas how I can speed this up? The pattern of points is
completely arbitrary and has no pattern in general.

James

2

Hi James,

···

On May 8, 2010, at 2:33 PM, James Sharpe wrote:

I'm trying to use H5S_select_elements to implement a parallel code that only reads in the elements that are in the current partition. This works fine when the number of elements in a partition is small but due to the linked list implementation of the point selection, destruction of the dataspace takes forever when I have on the order of 1 million+ point selections. Does anyone have any ideas how I can speed this up? The pattern of points is completely arbitrary and has no pattern in general.

  Hmm, we weren't really expecting anyone to use the element selections with that many points... :-/ Right now, we don't have any specific resources to apply to addressing this, but I'll file a bug in our issue tracker so it'll be in the general pool of things to work toward. As usual, high-quality patches would be welcome.

  Quincey