Hi Sebastian,

Thank you for the use case to replicate the problem. I managed to replicate and 
it is indeed a bug in the library caused by moving truncation of the file to 
its allocated EOA from the flush call to the close call, which someone did a 
while ago. I have entered Jira Bug HDFFV-9418 for this.

Thanks,
Mohamad

-----Original Message-----
From: Hdf-forum [mailto:[email protected]] On Behalf Of 
Sebastian Rettenberger
Sent: Monday, June 15, 2015 9:07 AM
To: HDF Users Discussion List
Subject: Re: [Hdf-forum] File state after flush and crash

Hi,

I think I figure out the problem:

The file size was not the real problem but the slightly different 
implementation for the large files.
To get a good performance on the parallel file system, I add some gaps in the 
dataset such that every task starts writing data at a multiple of the file 
system block size. This introduces "gaps" in the dataset with uninitialized 
values.
This itself is not a problem, however, the last task also added a gap at the 
end of the dataset which is never written. Thus, the file is smaller than 
expected.
An H5F_close() seems to fix either the header or the file size while a simple 
H5F_flush() does not.
Remove the gap from the last task solves the problem for me.

To reproduce this:
- Create a new file with a single dataset.
- Write parts of the dataset (make sure that some values at the end of the 
dataset are not initialized)
- Flush the file
- Crash the program
- Try to open the h5 file with h5dump or h5debug

I am using the MPIO backend and I have the 2 flags for the dataset:
H5Pset_layout(h5plist, H5D_CONTIGUOUS);
H5Pset_alloc_time(h5plist, H5D_ALLOC_TIME_EARLY); Not sure if this is important.

Let me know, if you still need a reproducer code.

Best regards,
Sebastian

On 06/10/2015 02:28 PM, Sebastian Rettenberger wrote:
> Hi,
>
> no, I do not modify the file that gets corrupted in do_work(). I only 
> access other HDF5 files.
>
> I figured out, that this problem only exists when creating large files 
> (> 3 TB) in parallel (> 1500 MPI tasks). For much smaller files, I did 
> not run into this problem.
>
> I will try to figure out the critical file size a create a replicator 
> but this might need some time since I have to wait for the compute 
> resources.
>
> I am not sure is this is helpful, but here is the error I get, when I 
> try to access the corrupt file with h5debug:
>
>> HDF5-DIAG: Error detected in HDF5 (1.8.11) thread 0:
>>   #000: H5F.c line 1582 in H5Fopen(): unable to open file
>>     major: File accessibilty
>>     minor: Unable to open file
>>   #001: H5F.c line 1373 in H5F_open(): unable to read superblock
>>     major: File accessibilty
>>     minor: Read failed
>>   #002: H5Fsuper.c line 351 in H5F_super_read(): unable to load 
>> superblock
>>     major: Object cache
>>     minor: Unable to protect metadata
>>   #003: H5AC.c line 1329 in H5AC_protect(): H5C_protect() failed.
>>     major: Object cache
>>     minor: Unable to protect metadata
>>   #004: H5C.c line 3570 in H5C_protect(): can't load entry
>>     major: Object cache
>>     minor: Unable to load metadata into cache
>>   #005: H5C.c line 7950 in H5C_load_entry(): unable to load entry
>>     major: Object cache
>>     minor: Unable to load metadata into cache
>>   #006: H5Fsuper_cache.c line 471 in H5F_sblock_load(): truncated
>> file: eof = 3968572377152, sblock->base_addr = 0, stored_eoa =
>> 3968574947328
>>     major: File accessibilty
>>     minor: File has been truncated
>> cannot open file
>
> Best regards,
> Sebastian
>
> On 06/08/2015 05:08 PM, Mohamad Chaarawi wrote:
>> Hi Sebastian,
>>
>> What happens in do_work()? Are you modifying the file in question?
>> If yes, then corruption can be expected..
>> If not, then the file should not be corrupted, and if it is then we 
>> have a bug in the library.
>>
>> If you can send a replicator for this problem we can investigate further.
>>
>> Thanks,
>> Mohamad
>>
>> -----Original Message-----
>> From: Hdf-forum [mailto:[email protected]] On 
>> Behalf Of Sebastian Rettenberger
>> Sent: Tuesday, June 02, 2015 4:27 AM
>> To: HDF Users Discussion List
>> Subject: [Hdf-forum] File state after flush and crash
>>
>> Hi,
>>
>> I want to save the program state of a parallel program in an HDF5 file.
>> For performance reason I do not want to open/close the file each time 
>> I write the program state, but use flush it instead.
>>
>> Thus my main loop looks basically like this:
>> while more_work() {
>>     do_work()
>>     update_hdf5_attributes()
>>     update hdf5_dataset()
>>     flush_hdf5()
>> }
>>
>> However, even if the program crashes during do_work(), I get a 
>> corrupt
>> HDF5 file.
>>
>> I found a short conversation regarding this but it is already 5 years
>> old:
>> http://lists.hdfgroup.org/pipermail/hdf-forum_lists.hdfgroup.org/2010
>> -February/002543.html
>>
>>
>> They mentioned that this might be a problem with the meta data cache.
>> Is this still true? Is there a way around it?
>> I also have some other open HDF5 file. Might this be a problem?
>>
>> Best regards,
>> Sebastian
>>
>> --
>> Sebastian Rettenberger, M.Sc.
>> Technische Universität München
>> Department of Informatics
>> Chair of Scientific Computing
>> Boltzmannstrasse 3, 85748 Garching, Germany http://www5.in.tum.de/
>>
>> _______________________________________________
>> Hdf-forum is for HDF software users discussion.
>> [email protected]
>> http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.o
>> rg
>> Twitter: https://twitter.com/hdf5
>>
>

--
Sebastian Rettenberger, M.Sc.
Technische Universität München
Department of Informatics
Chair of Scientific Computing
Boltzmannstrasse 3, 85748 Garching, Germany http://www5.in.tum.de/

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

Reply via email to