Quantcast

Parallel dataset resizing strategies

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

Parallel dataset resizing strategies

Chris Green

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.
We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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: Parallel dataset resizing strategies

Nelson, Jarom

If you can move to HDF5 1.10, I would recommend independent files for each MPI rank, and then create a master file (created independently perhaps by rank 0) with Virtual Datasets linking in the data from each rank in the format you need. Virtual Datasets can be created with file matching patterns for dynamically increasing datasets, so you might look into using that feature.

I found this approach much faster than creating a collective file (~5-10x speedup on a Lustre filesystem). You don’t need to do any collective reads or writes, and I think we could even bypass using parallel HDF5 altogether. Note, this will only work if you only ever need to open the Virtual Dataset in parallel (i.e. by more than one process) as non-collective read-only. If you need to have read-write access to the master file, you can’t access a Virtual Dataset using collective operations. You can, however, have as many processes as you like read from a virtual dataset from a file opened as read-only.

 

If you have other tools that use your data but can’t move to HDF5 1.10, you can h5repack a file with Virtual Datasets to remove the Virtual Datasets, and it should be compatible with HDF5 1.8 (use h5repack from HDF5 1.10 patch 1 or later). This also worked well for us and I was able to load a repacked file in IDL under a 1.8 HDF5 library. However h5repack is not a parallel application, so it can be slow to repack a very large file, on the order minutes per GB.

 

Jarom

 

From: Hdf-forum [mailto:[hidden email]] On Behalf Of Chris Green
Sent: Friday, July 22, 2016 9:32 AM
To: [hidden email]
Subject: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.

We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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: Parallel dataset resizing strategies

Chris Green

Hi,

Thanks for this. Comments inline.

On 7/22/16 12:13 PM, Nelson, Jarom wrote:

If you can move to HDF5 1.10, I would recommend independent files for each MPI rank, and then create a master file (created independently perhaps by rank 0) with Virtual Datasets linking in the data from each rank in the format you need. Virtual Datasets can be created with file matching patterns for dynamically increasing datasets, so you might look into using that feature.

We don't have existing tools relying on a particular version, so we are nominally free to move to HDF5 1.10.x. However, it won't be completely straightforward because I have been relying for now on using the homebrew version, which is currently 1.18.16. I'd have to dink the recipe to use 1.10.x, which is not a showstopper.

I found this approach much faster than creating a collective file (~5-10x speedup on a Lustre filesystem). You don’t need to do any collective reads or writes, and I think we could even bypass using parallel HDF5 altogether. Note, this will only work if you only ever need to open the Virtual Dataset in parallel (i.e. by more than one process) as non-collective read-only. If you need to have read-write access to the master file, you can’t access a Virtual Dataset using collective operations. You can, however, have as many processes as you like read from a virtual dataset from a file opened as read-only.

 

If you have other tools that use your data but can’t move to HDF5 1.10, you can h5repack a file with Virtual Datasets to remove the Virtual Datasets, and it should be compatible with HDF5 1.8 (use h5repack from HDF5 1.10 patch 1 or later). This also worked well for us and I was able to load a repacked file in IDL under a 1.8 HDF5 library. However h5repack is not a parallel application, so it can be slow to repack a very large file, on the order minutes per GB.

After having thought a little more about likely parallel models, I think now we can arrange that:

  • Only one rank will write to a particular dataset.

  • A dataset will not be read from in the same job in which it was written.

  • A dataset may be read by one or more ranks.

I *think* if that's the case, we could use a hierarchical multi-file format without resorting to virtual datasets, no? I still have some reading and experimenting to do, but if you have particular information that would speak to the likely success of this approach, I'd be happy to hear it.

Thanks,

Chris.

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Friday, July 22, 2016 9:32 AM
To: [hidden email]
Subject: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.

We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).


_______________________________________________
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

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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: Parallel dataset resizing strategies

Nelson, Jarom

If you want to have multiple ranks write to the same file, you’ll need to open the file in read-write and use parallel HDF5 with the associated overhead and complexity of the collective calls. I think the only way to avoid the overhead of the collective calls is to open separate files for each rank.

 

If you are going to have a multi-file approach, and read from files which are open in write mode by another process, you’ll need to have some way to get the metadata updated in the reading processes. It sounds like you might try another 1.10.x addition, the single-writer multiple-reader. If each rank can open its own output file in read-write, and all the other ranks’ files in read-only, you can avoid the parallel overhead. I haven’t tried this approach, and you’ll have to be careful of race conditions and keep the file metadata correct in all the ranks, but it sounds like it might fit your parallel I/O model. https://www.hdfgroup.org/HDF5/docNewFeatures/NewFeaturesSwmrDocs.html

 

Jarom

 

From: Hdf-forum [mailto:[hidden email]] On Behalf Of Chris Green
Sent: Monday, July 25, 2016 3:41 PM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

Thanks for this. Comments inline.

On 7/22/16 12:13 PM, Nelson, Jarom wrote:

If you can move to HDF5 1.10, I would recommend independent files for each MPI rank, and then create a master file (created independently perhaps by rank 0) with Virtual Datasets linking in the data from each rank in the format you need. Virtual Datasets can be created with file matching patterns for dynamically increasing datasets, so you might look into using that feature.

We don't have existing tools relying on a particular version, so we are nominally free to move to HDF5 1.10.x. However, it won't be completely straightforward because I have been relying for now on using the homebrew version, which is currently 1.18.16. I'd have to dink the recipe to use 1.10.x, which is not a showstopper.

I found this approach much faster than creating a collective file (~5-10x speedup on a Lustre filesystem). You don’t need to do any collective reads or writes, and I think we could even bypass using parallel HDF5 altogether. Note, this will only work if you only ever need to open the Virtual Dataset in parallel (i.e. by more than one process) as non-collective read-only. If you need to have read-write access to the master file, you can’t access a Virtual Dataset using collective operations. You can, however, have as many processes as you like read from a virtual dataset from a file opened as read-only.

 

If you have other tools that use your data but can’t move to HDF5 1.10, you can h5repack a file with Virtual Datasets to remove the Virtual Datasets, and it should be compatible with HDF5 1.8 (use h5repack from HDF5 1.10 patch 1 or later). This also worked well for us and I was able to load a repacked file in IDL under a 1.8 HDF5 library. However h5repack is not a parallel application, so it can be slow to repack a very large file, on the order minutes per GB.

After having thought a little more about likely parallel models, I think now we can arrange that:

·         Only one rank will write to a particular dataset.

·         A dataset will not be read from in the same job in which it was written.

·         A dataset may be read by one or more ranks.

I *think* if that's the case, we could use a hierarchical multi-file format without resorting to virtual datasets, no? I still have some reading and experimenting to do, but if you have particular information that would speak to the likely success of this approach, I'd be happy to hear it.

Thanks,

Chris.

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Friday, July 22, 2016 9:32 AM
To: [hidden email]
Subject: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.

We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).




_______________________________________________
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



-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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: Parallel dataset resizing strategies

Chris Green

Hi,

Thanks for continuing the conversation. I believe files will be *either* read from, *or* written to, but not both simultaneously, at least in the scenario I'm working on right now. I'd like to be able to write to the same file from different ranks simultaneously, but only to different datasets. If that's not possible without propagating dataset extension operations collectively to ranks not writing to that dataset, then I will start looking at the virtual dataset solution you suggested in your first reply.

Thanks again,
Chris.

On 7/25/16 7:41 PM, Nelson, Jarom wrote:

If you want to have multiple ranks write to the same file, you’ll need to open the file in read-write and use parallel HDF5 with the associated overhead and complexity of the collective calls. I think the only way to avoid the overhead of the collective calls is to open separate files for each rank.

 

If you are going to have a multi-file approach, and read from files which are open in write mode by another process, you’ll need to have some way to get the metadata updated in the reading processes. It sounds like you might try another 1.10.x addition, the single-writer multiple-reader. If each rank can open its own output file in read-write, and all the other ranks’ files in read-only, you can avoid the parallel overhead. I haven’t tried this approach, and you’ll have to be careful of race conditions and keep the file metadata correct in all the ranks, but it sounds like it might fit your parallel I/O model. https://www.hdfgroup.org/HDF5/docNewFeatures/NewFeaturesSwmrDocs.html

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Monday, July 25, 2016 3:41 PM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

Thanks for this. Comments inline.

On 7/22/16 12:13 PM, Nelson, Jarom wrote:

If you can move to HDF5 1.10, I would recommend independent files for each MPI rank, and then create a master file (created independently perhaps by rank 0) with Virtual Datasets linking in the data from each rank in the format you need. Virtual Datasets can be created with file matching patterns for dynamically increasing datasets, so you might look into using that feature.

We don't have existing tools relying on a particular version, so we are nominally free to move to HDF5 1.10.x. However, it won't be completely straightforward because I have been relying for now on using the homebrew version, which is currently 1.18.16. I'd have to dink the recipe to use 1.10.x, which is not a showstopper.

I found this approach much faster than creating a collective file (~5-10x speedup on a Lustre filesystem). You don’t need to do any collective reads or writes, and I think we could even bypass using parallel HDF5 altogether. Note, this will only work if you only ever need to open the Virtual Dataset in parallel (i.e. by more than one process) as non-collective read-only. If you need to have read-write access to the master file, you can’t access a Virtual Dataset using collective operations. You can, however, have as many processes as you like read from a virtual dataset from a file opened as read-only.

 

If you have other tools that use your data but can’t move to HDF5 1.10, you can h5repack a file with Virtual Datasets to remove the Virtual Datasets, and it should be compatible with HDF5 1.8 (use h5repack from HDF5 1.10 patch 1 or later). This also worked well for us and I was able to load a repacked file in IDL under a 1.8 HDF5 library. However h5repack is not a parallel application, so it can be slow to repack a very large file, on the order minutes per GB.

After having thought a little more about likely parallel models, I think now we can arrange that:

·         Only one rank will write to a particular dataset.

·         A dataset will not be read from in the same job in which it was written.

·         A dataset may be read by one or more ranks.

I *think* if that's the case, we could use a hierarchical multi-file format without resorting to virtual datasets, no? I still have some reading and experimenting to do, but if you have particular information that would speak to the likely success of this approach, I'd be happy to hear it.

Thanks,

Chris.

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Friday, July 22, 2016 9:32 AM
To: [hidden email]
Subject: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.

We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).




_______________________________________________
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



-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).


_______________________________________________
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

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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: Parallel dataset resizing strategies

Nelson, Jarom

Even if you are using independent datasets or groups from each process, the file metadata will need to be updated collectively for the HDF5 library to work correctly. Dataset identifiers, chunk locations, group information, attributes, etc. all need to be coordinated among all processes writing to the file for the file to have the correct data.

 

This page has a list of all the calls that need to be used collectively for the file to be written correctly.

https://www.hdfgroup.org/HDF5/doc/RM/CollectiveCalls.html

 

Jarom

 

From: Hdf-forum [mailto:[hidden email]] On Behalf Of Chris Green
Sent: Tuesday, July 26, 2016 8:33 AM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

Thanks for continuing the conversation. I believe files will be *either* read from, *or* written to, but not both simultaneously, at least in the scenario I'm working on right now. I'd like to be able to write to the same file from different ranks simultaneously, but only to different datasets. If that's not possible without propagating dataset extension operations collectively to ranks not writing to that dataset, then I will start looking at the virtual dataset solution you suggested in your first reply.

Thanks again,
Chris.

On 7/25/16 7:41 PM, Nelson, Jarom wrote:

If you want to have multiple ranks write to the same file, you’ll need to open the file in read-write and use parallel HDF5 with the associated overhead and complexity of the collective calls. I think the only way to avoid the overhead of the collective calls is to open separate files for each rank.

 

If you are going to have a multi-file approach, and read from files which are open in write mode by another process, you’ll need to have some way to get the metadata updated in the reading processes. It sounds like you might try another 1.10.x addition, the single-writer multiple-reader. If each rank can open its own output file in read-write, and all the other ranks’ files in read-only, you can avoid the parallel overhead. I haven’t tried this approach, and you’ll have to be careful of race conditions and keep the file metadata correct in all the ranks, but it sounds like it might fit your parallel I/O model. https://www.hdfgroup.org/HDF5/docNewFeatures/NewFeaturesSwmrDocs.html

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Monday, July 25, 2016 3:41 PM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

Thanks for this. Comments inline.

On 7/22/16 12:13 PM, Nelson, Jarom wrote:


If you can move to HDF5 1.10, I would recommend independent files for each MPI rank, and then create a master file (created independently perhaps by rank 0) with Virtual Datasets linking in the data from each rank in the format you need. Virtual Datasets can be created with file matching patterns for dynamically increasing datasets, so you might look into using that feature.

We don't have existing tools relying on a particular version, so we are nominally free to move to HDF5 1.10.x. However, it won't be completely straightforward because I have been relying for now on using the homebrew version, which is currently 1.18.16. I'd have to dink the recipe to use 1.10.x, which is not a showstopper.

I found this approach much faster than creating a collective file (~5-10x speedup on a Lustre filesystem). You don’t need to do any collective reads or writes, and I think we could even bypass using parallel HDF5 altogether. Note, this will only work if you only ever need to open the Virtual Dataset in parallel (i.e. by more than one process) as non-collective read-only. If you need to have read-write access to the master file, you can’t access a Virtual Dataset using collective operations. You can, however, have as many processes as you like read from a virtual dataset from a file opened as read-only.

 

If you have other tools that use your data but can’t move to HDF5 1.10, you can h5repack a file with Virtual Datasets to remove the Virtual Datasets, and it should be compatible with HDF5 1.8 (use h5repack from HDF5 1.10 patch 1 or later). This also worked well for us and I was able to load a repacked file in IDL under a 1.8 HDF5 library. However h5repack is not a parallel application, so it can be slow to repack a very large file, on the order minutes per GB.

After having thought a little more about likely parallel models, I think now we can arrange that:

·         Only one rank will write to a particular dataset.

·         A dataset will not be read from in the same job in which it was written.

·         A dataset may be read by one or more ranks.

I *think* if that's the case, we could use a hierarchical multi-file format without resorting to virtual datasets, no? I still have some reading and experimenting to do, but if you have particular information that would speak to the likely success of this approach, I'd be happy to hear it.

Thanks,

Chris.

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Friday, July 22, 2016 9:32 AM
To: [hidden email]
Subject: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.

We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).





_______________________________________________
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




-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).




_______________________________________________
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



-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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: Parallel dataset resizing strategies

Chris Green

Thanks for this, Jarom,

Based on some subsequent experiments, I think we're going with multiple files with a "master" file containing external links (not virtual, as a dataset is confined to a single file), and (possibly) an "extents" dataset describing where the data can be foun dfor each event in each data product. Life would be easier if I could send region references to the master file and have them "just work" with respect to the external links, but I haven't done the experiment yet (although I do have a forum post asking about this, see '"Re-seating" region references for external links?').

Thanks again for your posts -- they have helped us narrow in on a promising track.

Best,

Chris.


On 7/26/16 11:05 AM, Nelson, Jarom wrote:

Even if you are using independent datasets or groups from each process, the file metadata will need to be updated collectively for the HDF5 library to work correctly. Dataset identifiers, chunk locations, group information, attributes, etc. all need to be coordinated among all processes writing to the file for the file to have the correct data.

 

This page has a list of all the calls that need to be used collectively for the file to be written correctly.

https://www.hdfgroup.org/HDF5/doc/RM/CollectiveCalls.html

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Tuesday, July 26, 2016 8:33 AM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

Thanks for continuing the conversation. I believe files will be *either* read from, *or* written to, but not both simultaneously, at least in the scenario I'm working on right now. I'd like to be able to write to the same file from different ranks simultaneously, but only to different datasets. If that's not possible without propagating dataset extension operations collectively to ranks not writing to that dataset, then I will start looking at the virtual dataset solution you suggested in your first reply.

Thanks again,
Chris.

On 7/25/16 7:41 PM, Nelson, Jarom wrote:

If you want to have multiple ranks write to the same file, you’ll need to open the file in read-write and use parallel HDF5 with the associated overhead and complexity of the collective calls. I think the only way to avoid the overhead of the collective calls is to open separate files for each rank.

 

If you are going to have a multi-file approach, and read from files which are open in write mode by another process, you’ll need to have some way to get the metadata updated in the reading processes. It sounds like you might try another 1.10.x addition, the single-writer multiple-reader. If each rank can open its own output file in read-write, and all the other ranks’ files in read-only, you can avoid the parallel overhead. I haven’t tried this approach, and you’ll have to be careful of race conditions and keep the file metadata correct in all the ranks, but it sounds like it might fit your parallel I/O model. https://www.hdfgroup.org/HDF5/docNewFeatures/NewFeaturesSwmrDocs.html

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Monday, July 25, 2016 3:41 PM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

Thanks for this. Comments inline.

On 7/22/16 12:13 PM, Nelson, Jarom wrote:


If you can move to HDF5 1.10, I would recommend independent files for each MPI rank, and then create a master file (created independently perhaps by rank 0) with Virtual Datasets linking in the data from each rank in the format you need. Virtual Datasets can be created with file matching patterns for dynamically increasing datasets, so you might look into using that feature.

We don't have existing tools relying on a particular version, so we are nominally free to move to HDF5 1.10.x. However, it won't be completely straightforward because I have been relying for now on using the homebrew version, which is currently 1.18.16. I'd have to dink the recipe to use 1.10.x, which is not a showstopper.

I found this approach much faster than creating a collective file (~5-10x speedup on a Lustre filesystem). You don’t need to do any collective reads or writes, and I think we could even bypass using parallel HDF5 altogether. Note, this will only work if you only ever need to open the Virtual Dataset in parallel (i.e. by more than one process) as non-collective read-only. If you need to have read-write access to the master file, you can’t access a Virtual Dataset using collective operations. You can, however, have as many processes as you like read from a virtual dataset from a file opened as read-only.

 

If you have other tools that use your data but can’t move to HDF5 1.10, you can h5repack a file with Virtual Datasets to remove the Virtual Datasets, and it should be compatible with HDF5 1.8 (use h5repack from HDF5 1.10 patch 1 or later). This also worked well for us and I was able to load a repacked file in IDL under a 1.8 HDF5 library. However h5repack is not a parallel application, so it can be slow to repack a very large file, on the order minutes per GB.

After having thought a little more about likely parallel models, I think now we can arrange that:

·         Only one rank will write to a particular dataset.

·         A dataset will not be read from in the same job in which it was written.

·         A dataset may be read by one or more ranks.

I *think* if that's the case, we could use a hierarchical multi-file format without resorting to virtual datasets, no? I still have some reading and experimenting to do, but if you have particular information that would speak to the likely success of this approach, I'd be happy to hear it.

Thanks,

Chris.

 

Jarom

 

From: Hdf-forum [[hidden email]] On Behalf Of Chris Green
Sent: Friday, July 22, 2016 9:32 AM
To: [hidden email]
Subject: [Hdf-forum] Parallel dataset resizing strategies

 

Hi,

I am relatively new to HDF5 and HDF5/parallel, and although I have experience with MPI it is not extensive. We are exploring ways of saving data in parallel using HDF5 in a field in which it is practically unknown up to now.

Our paradigm is "parallel modular event processing:"

  • A typical job processes many "events."
  • An event contains all of the interesting data (raw and processed) associated with some time interval.
  • Each event can be processed independently of all other events.
  • Each event's data can be subdivided into internal components, "data products."
  • "Modules" are processing subunits which read or generate one or more data products for each event.
  • One can calculate a data dependency graph specifying the allowed ordering and/or parallelism of modules processing one or more events simultaneously for a given job configuration and event structure.

We have been using h5py with HDF5 and OpenMPI to explore different strategies for parallel I/O in a future parallel event-processing framework. One of the approaches we have come up with so far is to have one HDF5 dataset per unique data product / writer module combination, keeping track of the different relevant sections of each dataset via (for now) an external database. This works well in serial tests, but in parallel tests we are running up against the constraint that dataset resizing is a collective operation, meaning that all ranks including non-writers will have to become aware of and duplicate dataset resizing operations required by other writers. The problem seems to get even worse if there's a possibility that two or more instances of a module would need to extend and write to the same dataset at the same time (while processing different events, say), since they will have to coordinate and agree on the new size of the dataset and their respective sections thereof.

Are we misunderstanding the problem, or is it really this hard? Has anyone else hit upon a reasonable strategy for handling this or something like it?

Any pointers appreciated.

Thanks,

Chris Green.

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).





_______________________________________________
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




-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).




_______________________________________________
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



-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).


_______________________________________________
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

-- 
Chris Green [hidden email], FNAL CS/SCD/ADSS/SSI/TAC;
'phone (630) 840-2167; Skype: chris.h.green;
IM: [hidden email], chissgreen (AIM),
chris.h.green (Google Talk).

_______________________________________________
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...