Collective H5Dopen

classic Classic list List threaded Threaded
7 messages Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Collective H5Dopen

Lion Krischer
Hi all,

the HDF5 documentation states that H5Dopen may be called independently
if the target object is not modified. What exactly is the definition of
a modification in that case? Does writing to the data set count or is it
only resizing/attribute setting/...

Our application has to write in parallel to on the order order of 1
million independent data sets in a single file. We currently do it by
creating the file up-front and closing it again. Each rank then opens
only the data sets he wants to write to and does its thing. Note that at
this point only the actual data values change - everything else is
already in the file.

At least this is what we would like to do - currently we have to call
H5Dopen for all datasets on each rank to get it to work correctly.
Interestingly enough it also appears to work if each rank opens exactly
the same number of data sets.

Can H5Dopen() work in this way?

If not: can you advise us on the fastest way of writing 1 million
separate data sets in parallel on many cores?

All the best and thanks!

Lion


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Collective H5Dopen

Miller, Mark C.

 

 

"Hdf-forum on behalf of Lion Krischer" wrote:

 

Hi all,

 

the HDF5 documentation states that H5Dopen may be called independently

if the target object is not modified. What exactly is the definition of

a modification in that case? Does writing to the data set count or is it

only resizing/attribute setting/...

 

Hmm. Interesting question. I would think *any* operation that changes data (either

metadata or raw data) is a *modification*. But, I think I kinda see where you are

going here...If you've created a (non-extendible) dataset with no checksum or compression

filters etc., all you wanna do is change the raw data but not perterb any of the HDF5 file's

metadata. I don't think the HDF5 library would treat that as a non-modification though. I think a write

operation (even on the raw data) can wind up changing how the library caches dataset and file

metadata in memory thereby creating a situation where two different tasks have a different idea

of the file's cached metadata. When it comes time to close the file, which tasks' view of the metadata

is correct?

 

Its concievable you could manually do it by obtaining the dataset's offset in the file, calling

any necessary H5Tconvert method on your buffer just prior to writing in and then writing the

buffer yourself to the file via pwrite or something. That would essentially skirt HDF5 though

and probably be too complex to be worth it.

 

It might be worth looking at, H5DOwrite_chunk() though, https://support.hdfgroup.org/HDF5/doc/HL/RM_HDF5Optimized.html

to ask how that works and if it achieves or gets close to behavior you want.

 

Mark

 

 

 

 

_______________________________________________

Hdf-forum is for HDF software users discussion.

 


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Collective H5Dopen

Lion Krischer

Thanks a lot for your input!

We'll try to directly use MPI I/O to write to the file after getting the offsets from the initially created file. Thanks for the suggestion! As we use none of the advanced features a straight mmap or its equivalent with MPI I/O should work.

Our code guarantees that each data set is only written to from a single rank so that should work and metadata synchronization issues do not really exist. It would be great if something like this could be supported directly via the hdf5 library. Maybe a H5DOwrite_directly() that does some very basic checks and then writes straight to the uncompressed, unchunked, .. dataset without even touching the metadata so it should work fine with parallel I/O?

Does H5DOwrite_chunk() work for parallel code? I cannot really tell from the documentation and it also appears to solve the way more complicated problem of efficiently writing chunked data.

We'd still be happy to hear additional suggestions :)


Cheers!

Lion


Hmm. Interesting question. I would think *any* operation that changes data (either

metadata or raw data) is a *modification*. But, I think I kinda see where you are

going here...If you've created a (non-extendible) dataset with no checksum or compression

filters etc., all you wanna do is change the raw data but not perterb any of the HDF5 file's

metadata. I don't think the HDF5 library would treat that as a non-modification though. I think a write

operation (even on the raw data) can wind up changing how the library caches dataset and file

metadata in memory thereby creating a situation where two different tasks have a different idea

of the file's cached metadata. When it comes time to close the file, which tasks' view of the metadata

is correct?

 

Its concievable you could manually do it by obtaining the dataset's offset in the file, calling

any necessary H5Tconvert method on your buffer just prior to writing in and then writing the

buffer yourself to the file via pwrite or something. That would essentially skirt HDF5 though

and probably be too complex to be worth it.

 

It might be worth looking at, H5DOwrite_chunk() though, https://support.hdfgroup.org/HDF5/doc/HL/RM_HDF5Optimized.html

to ask how that works and if it achieves or gets close to behavior you want.

 

Mark

 

 

 

 

_______________________________________________

Hdf-forum is for HDF software users discussion.

 



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Collective H5Dopen

bljones

Hi Lion,

 

I checked with the developers…

H5DOwrite_chunk() does not work with parallel HDF5.

 

-Barbara

[hidden email]

 

From: Hdf-forum [mailto:[hidden email]] On Behalf Of Lion Krischer
Sent: Tuesday, March 14, 2017 3:43 PM
To: [hidden email]
Subject: Re: [Hdf-forum] Collective H5Dopen

 

Thanks a lot for your input!

We'll try to directly use MPI I/O to write to the file after getting the offsets from the initially created file. Thanks for the suggestion! As we use none of the advanced features a straight mmap or its equivalent with MPI I/O should work.

Our code guarantees that each data set is only written to from a single rank so that should work and metadata synchronization issues do not really exist. It would be great if something like this could be supported directly via the hdf5 library. Maybe a H5DOwrite_directly() that does some very basic checks and then writes straight to the uncompressed, unchunked, .. dataset without even touching the metadata so it should work fine with parallel I/O?

Does H5DOwrite_chunk() work for parallel code? I cannot really tell from the documentation and it also appears to solve the way more complicated problem of efficiently writing chunked data.

We'd still be happy to hear additional suggestions :)


Cheers!

Lion



Hmm. Interesting question. I would think *any* operation that changes data (either

metadata or raw data) is a *modification*. But, I think I kinda see where you are

going here...If you've created a (non-extendible) dataset with no checksum or compression

filters etc., all you wanna do is change the raw data but not perterb any of the HDF5 file's

metadata. I don't think the HDF5 library would treat that as a non-modification though. I think a write

operation (even on the raw data) can wind up changing how the library caches dataset and file

metadata in memory thereby creating a situation where two different tasks have a different idea

of the file's cached metadata. When it comes time to close the file, which tasks' view of the metadata

is correct?

 

Its concievable you could manually do it by obtaining the dataset's offset in the file, calling

any necessary H5Tconvert method on your buffer just prior to writing in and then writing the

buffer yourself to the file via pwrite or something. That would essentially skirt HDF5 though

and probably be too complex to be worth it.

 

It might be worth looking at, H5DOwrite_chunk() though, https://support.hdfgroup.org/HDF5/doc/HL/RM_HDF5Optimized.html

to ask how that works and if it achieves or gets close to behavior you want.

 

Mark

 

 

 

 

_______________________________________________

Hdf-forum is for HDF software users discussion.

 




_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5

 


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Collective H5Dopen

Lion Krischer

Hi Barbara and Mark,

thanks for getting back to me.

As we tried to implement direct MPI I/O we discovered that HDF5 does not directly allocate created datasets so we could not get the offsets to do the direct writing. Doing

auto plist = H5Pcreate(H5P_DATASET_CREATE);
H5Pset_layout(plist, H5D_CONTIGUOUS);
H5Pset_alloc_time(plist, H5D_ALLOC_TIME_EARLY);

before calling H5Dcreate(); and passing plist does allocate the data - then the offsets works.

Turns out that closing the serially written file, reopening it in parallel and writing independently with H5DOpen() and H5Dwrite() then works fine. So I guess the modification of the dataset does then not count as a modification and it does not have to be collective. The files close cleanly and the output is correct.

Hope it helps someone who runs into the same issue!

It would also be very helpful to document what exactly counts as a modification the collective/independent call page.

Cheers!

Lion


On 20/03/2017 18:13, Barbara Jones wrote:

Hi Lion,

 

I checked with the developers…

H5DOwrite_chunk() does not work with parallel HDF5.

 

-Barbara

[hidden email]

 

From: Hdf-forum [[hidden email]] On Behalf Of Lion Krischer
Sent: Tuesday, March 14, 2017 3:43 PM
To: [hidden email]
Subject: Re: [Hdf-forum] Collective H5Dopen

 

Thanks a lot for your input!

We'll try to directly use MPI I/O to write to the file after getting the offsets from the initially created file. Thanks for the suggestion! As we use none of the advanced features a straight mmap or its equivalent with MPI I/O should work.

Our code guarantees that each data set is only written to from a single rank so that should work and metadata synchronization issues do not really exist. It would be great if something like this could be supported directly via the hdf5 library. Maybe a H5DOwrite_directly() that does some very basic checks and then writes straight to the uncompressed, unchunked, .. dataset without even touching the metadata so it should work fine with parallel I/O?

Does H5DOwrite_chunk() work for parallel code? I cannot really tell from the documentation and it also appears to solve the way more complicated problem of efficiently writing chunked data.

We'd still be happy to hear additional suggestions :)


Cheers!

Lion



Hmm. Interesting question. I would think *any* operation that changes data (either

metadata or raw data) is a *modification*. But, I think I kinda see where you are

going here...If you've created a (non-extendible) dataset with no checksum or compression

filters etc., all you wanna do is change the raw data but not perterb any of the HDF5 file's

metadata. I don't think the HDF5 library would treat that as a non-modification though. I think a write

operation (even on the raw data) can wind up changing how the library caches dataset and file

metadata in memory thereby creating a situation where two different tasks have a different idea

of the file's cached metadata. When it comes time to close the file, which tasks' view of the metadata

is correct?

 

Its concievable you could manually do it by obtaining the dataset's offset in the file, calling

any necessary H5Tconvert method on your buffer just prior to writing in and then writing the

buffer yourself to the file via pwrite or something. That would essentially skirt HDF5 though

and probably be too complex to be worth it.

 

It might be worth looking at, H5DOwrite_chunk() though, https://support.hdfgroup.org/HDF5/doc/HL/RM_HDF5Optimized.html

to ask how that works and if it achieves or gets close to behavior you want.

 

Mark

 

 

 

 

_______________________________________________

Hdf-forum is for HDF software users discussion.

 




_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5

 



_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Collective H5Dopen

Miller, Mark C.

Hi Lion,

 

My apologies if my guesses about HDF5 behavior were misleading.

 

I had hoped that if I stated something patently wrong about HDF5, a more cognizant developer might speak up and correct me ;)

 

Glad you got things working and thanks for this information.

 

Mark

 

 

"Hdf-forum on behalf of Lion Krischer" wrote:

 

Hi Barbara and Mark,

thanks for getting back to me.

As we tried to implement direct MPI I/O we discovered that HDF5 does not directly allocate created datasets so we could not get the offsets to do the direct writing. Doing

auto plist = H5Pcreate(H5P_DATASET_CREATE);
H5Pset_layout(plist, H5D_CONTIGUOUS);
H5Pset_alloc_time(plist, H5D_ALLOC_TIME_EARLY);

before calling H5Dcreate(); and passing plist does allocate the data - then the offsets works.

Turns out that closing the serially written file, reopening it in parallel and writing independently with H5DOpen() and H5Dwrite() then works fine. So I guess the modification of the dataset does then not count as a modification and it does not have to be collective. The files close cleanly and the output is correct.

Hope it helps someone who runs into the same issue!

It would also be very helpful to document what exactly counts as a modification the collective/independent call page.

Cheers!

Lion

 

On 20/03/2017 18:13, Barbara Jones wrote:

Hi Lion,

 

I checked with the developers…

H5DOwrite_chunk() does not work with parallel HDF5.

 

-Barbara

[hidden email]

 

From: Hdf-forum [[hidden email]] On Behalf Of Lion Krischer
Sent: Tuesday, March 14, 2017 3:43 PM
To: [hidden email]
Subject: Re: [Hdf-forum] Collective H5Dopen

 

Thanks a lot for your input!

We'll try to directly use MPI I/O to write to the file after getting the offsets from the initially created file. Thanks for the suggestion! As we use none of the advanced features a straight mmap or its equivalent with MPI I/O should work.

Our code guarantees that each data set is only written to from a single rank so that should work and metadata synchronization issues do not really exist. It would be great if something like this could be supported directly via the hdf5 library. Maybe a H5DOwrite_directly() that does some very basic checks and then writes straight to the uncompressed, unchunked, .. dataset without even touching the metadata so it should work fine with parallel I/O?

Does H5DOwrite_chunk() work for parallel code? I cannot really tell from the documentation and it also appears to solve the way more complicated problem of efficiently writing chunked data.

We'd still be happy to hear additional suggestions :)


Cheers!

Lion




Hmm. Interesting question. I would think *any* operation that changes data (either

metadata or raw data) is a *modification*. But, I think I kinda see where you are

going here...If you've created a (non-extendible) dataset with no checksum or compression

filters etc., all you wanna do is change the raw data but not perterb any of the HDF5 file's

metadata. I don't think the HDF5 library would treat that as a non-modification though. I think a write

operation (even on the raw data) can wind up changing how the library caches dataset and file

metadata in memory thereby creating a situation where two different tasks have a different idea

of the file's cached metadata. When it comes time to close the file, which tasks' view of the metadata

is correct?

 

Its concievable you could manually do it by obtaining the dataset's offset in the file, calling

any necessary H5Tconvert method on your buffer just prior to writing in and then writing the

buffer yourself to the file via pwrite or something. That would essentially skirt HDF5 though

and probably be too complex to be worth it.

 

It might be worth looking at, H5DOwrite_chunk() though, https://support.hdfgroup.org/HDF5/doc/HL/RM_HDF5Optimized.html

to ask how that works and if it achieves or gets close to behavior you want.

 

Mark

 

 

 

 

_______________________________________________

Hdf-forum is for HDF software users discussion.

 





_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5

 




_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5




_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Collective H5Dopen

Lion Krischer

Hi Mark,

no need to be sorry - your hint did lead to a working solution after all :)

Cheers!

Lion



On 20/03/2017 20:44, Miller, Mark C. wrote:

Hi Lion,

 

My apologies if my guesses about HDF5 behavior were misleading.

 

I had hoped that if I stated something patently wrong about HDF5, a more cognizant developer might speak up and correct me ;)

 

Glad you got things working and thanks for this information.

 

Mark

 

 

"Hdf-forum on behalf of Lion Krischer" wrote:

 

Hi Barbara and Mark,

thanks for getting back to me.

As we tried to implement direct MPI I/O we discovered that HDF5 does not directly allocate created datasets so we could not get the offsets to do the direct writing. Doing

auto plist = H5Pcreate(H5P_DATASET_CREATE);
H5Pset_layout(plist, H5D_CONTIGUOUS);
H5Pset_alloc_time(plist, H5D_ALLOC_TIME_EARLY);

before calling H5Dcreate(); and passing plist does allocate the data - then the offsets works.

Turns out that closing the serially written file, reopening it in parallel and writing independently with H5DOpen() and H5Dwrite() then works fine. So I guess the modification of the dataset does then not count as a modification and it does not have to be collective. The files close cleanly and the output is correct.

Hope it helps someone who runs into the same issue!

It would also be very helpful to document what exactly counts as a modification the collective/independent call page.

Cheers!

Lion

 

On 20/03/2017 18:13, Barbara Jones wrote:

Hi Lion,

 

I checked with the developers…

H5DOwrite_chunk() does not work with parallel HDF5.

 

-Barbara

[hidden email]

 

From: Hdf-forum [[hidden email]] On Behalf Of Lion Krischer
Sent: Tuesday, March 14, 2017 3:43 PM
To: [hidden email]
Subject: Re: [Hdf-forum] Collective H5Dopen

 

Thanks a lot for your input!

We'll try to directly use MPI I/O to write to the file after getting the offsets from the initially created file. Thanks for the suggestion! As we use none of the advanced features a straight mmap or its equivalent with MPI I/O should work.

Our code guarantees that each data set is only written to from a single rank so that should work and metadata synchronization issues do not really exist. It would be great if something like this could be supported directly via the hdf5 library. Maybe a H5DOwrite_directly() that does some very basic checks and then writes straight to the uncompressed, unchunked, .. dataset without even touching the metadata so it should work fine with parallel I/O?

Does H5DOwrite_chunk() work for parallel code? I cannot really tell from the documentation and it also appears to solve the way more complicated problem of efficiently writing chunked data.

We'd still be happy to hear additional suggestions :)


Cheers!

Lion




Hmm. Interesting question. I would think *any* operation that changes data (either

metadata or raw data) is a *modification*. But, I think I kinda see where you are

going here...If you've created a (non-extendible) dataset with no checksum or compression

filters etc., all you wanna do is change the raw data but not perterb any of the HDF5 file's

metadata. I don't think the HDF5 library would treat that as a non-modification though. I think a write

operation (even on the raw data) can wind up changing how the library caches dataset and file

metadata in memory thereby creating a situation where two different tasks have a different idea

of the file's cached metadata. When it comes time to close the file, which tasks' view of the metadata

is correct?

 

Its concievable you could manually do it by obtaining the dataset's offset in the file, calling

any necessary H5Tconvert method on your buffer just prior to writing in and then writing the

buffer yourself to the file via pwrite or something. That would essentially skirt HDF5 though

and probably be too complex to be worth it.

 

It might be worth looking at, H5DOwrite_chunk() though, https://support.hdfgroup.org/HDF5/doc/HL/RM_HDF5Optimized.html

to ask how that works and if it achieves or gets close to behavior you want.

 

Mark

 

 

 

 

_______________________________________________

Hdf-forum is for HDF software users discussion.

 





_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5

 




_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5





_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5


_______________________________________________
Hdf-forum is for HDF software users discussion.
[hidden email]
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
Loading...