When you parallel-hdf5, do all attribute and group creations need to be called collectively by all ranks?
Specifically for attributes, do all ranks have to participate in the H5Awrite, or does only one rank write the data?
Lastly, if one has a series of H5Awrites, can those be packaged together so that there are a bunch of small writes being done?
thanks,
kevin harms
Hi Kevin,
When you parallel-hdf5, do all attribute and group creations need to be called collectively by all ranks?
Yes. All metadata modifications need to be collective operations.
Specifically for attributes, do all ranks have to participate in the H5Awrite, or does only one rank write the data?
Currently, yes.
Lastly, if one has a series of H5Awrites, can those be packaged together so that there are a bunch of small writes being done?
Not at this time.
Quincey
On Jul 29, 2010, at 3:36 PM, Kevin Harms wrote:
Quincey,
The H5Awrite only results in one rank writing the data though? Would you have any suggestions on how to speed up setting a number of attributes?
thanks,
kevin harms
On Jul 29, 2010, at 4:04 PM, Quincey Koziol wrote:
Hi Kevin,
On Jul 29, 2010, at 3:36 PM, Kevin Harms wrote:
When you parallel-hdf5, do all attribute and group creations need to be called collectively by all ranks?
Yes. All metadata modifications need to be collective operations.
Specifically for attributes, do all ranks have to participate in the H5Awrite, or does only one rank write the data?
Currently, yes.
Lastly, if one has a series of H5Awrites, can those be packaged together so that there are a bunch of small writes being done?
Not at this time.
Quincey
_______________________________________________
Hdf-forum is for HDF software users discussion.
Hdf-forum@hdfgroup.org
http://mail.hdfgroup.org/mailman/listinfo/hdf-forum_hdfgroup.org
Hi Kevin,
Quincey,
The H5Awrite only results in one rank writing the data though?
Eventually, yes, but which rank will do it can't be predicted. (The metadata writes are divided among all the processes and rely on all the metadata operations on all the processes being identical)
Would you have any suggestions on how to speed up setting a number of attributes?
Can you operate on the file in sequential mode? Possibly before/after operating on it in parallel?
Quincey
On Jul 29, 2010, at 4:38 PM, Kevin Harms wrote:
thanks,
kevin harmsOn Jul 29, 2010, at 4:04 PM, Quincey Koziol wrote:
Hi Kevin,
On Jul 29, 2010, at 3:36 PM, Kevin Harms wrote:
When you parallel-hdf5, do all attribute and group creations need to be called collectively by all ranks?
Yes. All metadata modifications need to be collective operations.
Specifically for attributes, do all ranks have to participate in the H5Awrite, or does only one rank write the data?
Currently, yes.
Lastly, if one has a series of H5Awrites, can those be packaged together so that there are a bunch of small writes being done?
Not at this time.
Quincey
_______________________________________________
Hdf-forum is for HDF software users discussion.
Hdf-forum@hdfgroup.org
http://mail.hdfgroup.org/mailman/listinfo/hdf-forum_hdfgroup.org_______________________________________________
Hdf-forum is for HDF software users discussion.
Hdf-forum@hdfgroup.org
http://mail.hdfgroup.org/mailman/listinfo/hdf-forum_hdfgroup.org
