Hi all:
The fact that HDF5 is not thread-safe is horrible actually and should be
fixed some time. The current model is more compatible with the 90s
computing models, not the current one. Actually I learned that HDF5 is not
thread-safe the hard way, and it ruined many of my plans for the use of the
library. I don't like/trust having build options that make my program
works, and I believe more in separating source writing from building, which
is why I decided I'm not going to use the --enable-threadsafe option, and
decided to use my own locks with std::thread (or boost::thread if you don't
use C++11). I just wrapped every H5 call with a
std::unique_lock<std::mutex>, and that solved the problem. For me this
wasn't a big change, because I use C++ classes and templates and
compile-time objects and I have a class with all the H5 functions I needed,
which were not too many. It took me like an hour to fix my program, and
it's working now.
Just wanted to share my experience.
And yes, I also think requiring certain build options is a little iffy. But
in my case, the deployment platform is Ubuntu, where HDF5 is built with
this option. So if I were to be prudent and do my own locking even when
using the C API, I would pay for double locking on my main (so far only)
target platform.
Elvis
Cheers,
Samer Afach>
>>>
>>> There was some recent discussion that only the C API of HDF5 is
>>> threadsafe,
>>> but not the C++ layer on top of it. To be safe you should probably
keep
>>> at
>>> the C API.
>>
>> Aha. Thanks for the heads up Werner.
>>
>> The FAQ has this to say:
>>
>> "By default, you cannot build either Parallel HDF5 with C++ or
>> Parallel HDF5 with the thread-safe feature. You will receive a
>> configure error if you try either of these combinations."
>>
>> But it does not say that C++ with the thread-safe is unsupported
(when
>> using serial HDF5). Maybe the FAQ should be updated?
>>
>> This is quite unfortunate since our app is C++ and the C++ API is so
>> much more convenient, but I guess I'll port our code to the C API
>> (it's not that much code).
>
> What you probably could do as well is to have your own lock with one
global
> mutex around each call to the C++ API. HDF5 itself allows only one
running
> thread internally as well and just puts such a lock inside the API,
so it
> would come to the same.
This is what I ended up doing, so thanks a lot for the suggestion.
It was much easier than I thought because, since I was using Qt
anyway, I could just create a global QMutex, and then create a
QMutexLocker (RAII-style mutex locker) on the stack in each function
(just two of them) where I use HDF5.This sort of coarse locking is OK for me since overwhelming majority
of time is spent in a HDF5 H5::DataSpace::read(..) call (I'm reading
big compressed data). The other HDF5 calls are very quick by
comparison.Just a final question about this. The FAQ states that the threadsafe
guarantee does not extent to the high-level API.
I just want to confirm that this is really the case: Is the threadsafe
guarantee only for the low-level C API?
The reason I'm asking is that I'm now considering porting my HDF5 usage
from C++ to C, to avoid having to do the manual locking myself, and the
high-level API looked quite convenient for what I want to do. But if it's
really not threadsafe, I'll stick to the regular C API.
Nevermind, I found a very definitive answer in the 1.8.16 release notes,
where the --enable-threadsafe --enable-hl combination was marked as
unsupported in the build system:
- The thread-safety + high-level library combination has been marked
as "unsupported" in the AutotoolsThe global lock used by the thread-safety feature has never been
raised to the high-level library level, making it possible that
the library state could change if a context switch were to occur in
a high-level library call. Because of this, the combination of
thread-safety and high-level library is officially unsupported by
The HDF Group.In the past, although this combination has never been supported,
this
was not enforced by the build systems. These changes will cause an
Autotools configure step to fail if --enable-threadsafe and
--enable-hl are combined unless additional options are specified.
Since the high-level library is built by default, this means that
these extra configuration options will need to be used any time
--enable-threadsafe is selected.To build with --enable-threadsafe, either:
1) Use --disable-hl to disable the high-level library (recommended)
2) Use --enable-unsupported to build the high-level library with
the thread-safety feature.So I'll stick to the low-level C API.
Elvis
Thanks in advance,
ElvisElvis
>
> Werner
>
>
>
>> So far we haven't had any problems under Ubuntu (where the package
is
>> built with --enable-threadsafe --enable-unsupported), but I guess
>> we've just been lucky.
>>
>> Elvis
>>
>>> Cheers,
>>>
>>> Werner
>>>
>>>
>>>
>>>>
>>>>>
>>>>> Yes it is still the case that you cannot enable C++ or Fortran
(or the
>>>>> High Level APIs) when threadsafe is enabled. —enable-unsupported
can
>>>>> override this behavior.
>>>>
>>>> Aha, so that's why I had to enable that option
>>>>
>>>> I see now that this is what the Ubuntu package does. I'll ask the
Arch
>>>> package maintainer if he's willing to do the same.
>>>>
>>>> I've confirmed now that I don't have any problems with my program
if I
>>>> re-built the Arch package with --enable-threadsafe
>>>> --enable-unsupported.
>>>>
>>>> Thanks for the info!
>>>>
>>>> Elvis
>>>>
>>>>> Scot
>>>>>
>>>>>
>>>>>>
>>>>>> <elvis.stansvik@orexplore.com>:
>>>>>>>
>>>>>>>>
>>>>>>>> Hi Elvis,
>>>>>>>>
>>>>>>>> Did you build your HDF5 library with thread-safety enabled
>>>>>>>> (--enable-threadsafe w/ configure)?
>>>>>>>
>>>>>>> Hi Dana, and thanks for the quick reply. I think we just
e-mailed
>>>>>>> past
>>>>>>> each other (see my previous mail).
>>>>>>>
>>>>>>> I wrongly called it --thread-safe in that mail, but it was
>>>>>>> --enable-threadsafe I was referring to. But yes, I'm pretty
sure this
>>>>>>> is the problem.
>>>>>>>
>>>>>>> I'm rebuilding the Arch package now with --enable-threadsafe.
>>>>>>
>>>>>> I spoke a little too soon. I now found this bug filed against
the Arch
>>>>>> package:
>>>>>>
>>>>>> https://bugs.archlinux.org/task/33805
>>>>>>
>>>>>> The reporter asked the package maintainer to add
--enable-threadsafe,
>>>>>> but the package maintainer closed the bug saying that
>>>>>> --enable-threadsafe is not compatible with the Fortran build
(in Arch,
>>>>>> the C++ and Fortran APIs are bundled into one package
>>>>>> hdf5-cpp-fortran).
>>>>>>
>>>>>> Anyone know if that is still the case? If so I can't open a bug
>>>>>> against the package again asking for --enable-threadsafe to be
added.
>>>>>> But I could open a bug asking the package to be split I guess.
>>>>>>
>>>>>> Elvis
>>>>>>
>>>>>>> Elvis
>>>>>>>
>>>>>>>> Dana Robinson
>>>>>>>> Software Engineer
>>>>>>>> The HDF Group
>>>>>>>>
>>>>>>>> Get Outlook for Android
>>>>>>>>
>>>>>>>> From: Elvis Stansvik
>>>>>>>> Sent: Thursday, September 22, 12:43
>>>>>>>> Subject: [Hdf-forum] Simply using the library from separate
threads
>>>>>>>> (C++
>>>>>>>> API)
>>>>>>>> To: HDF Users Discussion List
>>>>>>>>
>>>>>>>> Hi all, I'm using the C++ API to read HDF5 files from separate
>>>>>>>> threads
>>>>>>>> (no
>>>>>>>> writing). None of my threads read the same file, but they do
execute
>>>>>>>> simultaneously. The reason I'm using threading is not to speed
>>>>>>>> things
>>>>>>>> up or
>>>>>>>> get better throughput, but simply to not block the UI (it's Qt
>>>>>>>> application)
>>>>>>>> while reading. So this is not about "Parallel HDF5" or
anything,
>>>>>>>> just
>>>>>>>> simply
>>>>>>>> using the serial library "from scratch" from multiple
threads. This
>>>>>>>> has been
>>>>>>>> working fine when testing on Ubuntu 16.04 (our target OS),
which has
>>>>>>>> HDF5
>>>>>>>> 1.8.16. I recently tested on my personal Arch Linux machine
though,
>>>>>>>> which
>>>>>>>> has HDF5 1.10.0, and got this segmentation fault: (gdb) bt #0
>>>>>>>> 0x00007ffff67c57d9 in H5SL_search () from
/usr/lib/libhdf5.so.100 #1
>>>>>>>> 0x00007ffff678dd19 in H5P_copy_plist () from
/usr/lib/libhdf5.so.100
>>>>>>>> #2
>>>>>>>> 0x00007ffff66a7fc0 in H5F_new () from /usr/lib/libhdf5.so.100
#3
>>>>>>>> 0x00007ffff66a8f55 in H5F_open () from
/usr/lib/libhdf5.so.100 #4
>>>>>>>> 0x00007ffff66a155d in H5Fopen () from /usr/lib/libhdf5.so.100
#5
>>>>>>>> 0x00007ffff6b79546 in H5::H5File::p_get_file(char const*,
unsigned
>>>>>>>> int,
>>>>>>>> H5::FileCreatPropList const&, H5::FileAccPropList const&) ()
from
>>>>>>>> /usr/lib/libhdf5_cpp.so.100 #6 0x00007ffff6b79750 in
>>>>>>>> H5::H5File::H5File(char
>>>>>>>> const*, unsigned int, H5::FileCreatPropList const&,
>>>>>>>> H5::FileAccPropList
>>>>>>>> const&) () from /usr/lib/libhdf5_cpp.so.100 #7
0x000000000041f00e in
>>>>>>>> HDF5ImageReader::RequestInformation (this=0x7fffbc002de0,
>>>>>>>> request=0x7fffbc010da0, inputVector=0x0,
>>>>>>>> outputVector=0x7fffbc0039d0)
>>>>>>>> at
>>>>>>>>
>>>>>>>>
>>>>>>>>
/home/estan/Projekt/orexplore/dev/src/insight/src/model/HDF5ImageReader.cpp:91
>>>>>>>> #8 0x00007fffee8200d0 in
>>>>>>>> vtkExecutive::CallAlgorithm(vtkInformation*,
>>>>>>>> int,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #9
0x00007fffee837fa9 in
>>>>>>>>
>>>>>>>>
vtkStreamingDemandDrivenPipeline::ExecuteInformation(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #10
0x00007fffee81ce05 in
>>>>>>>> vtkDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #11
0x00007fffee835c55 in
>>>>>>>>
vtkStreamingDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #12
0x00007fffee816e1a in
>>>>>>>> vtkCompositeDataPipeline::ForwardUpstream(vtkInformation*) ()
from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #13
0x00007fffee81ccb5 in
>>>>>>>> vtkDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #14
0x00007fffee835c55 in
>>>>>>>>
vtkStreamingDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #15
0x00007fffee816e1a in
>>>>>>>> vtkCompositeDataPipeline::ForwardUpstream(vtkInformation*) ()
from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #16
0x00007fffee81ccb5 in
>>>>>>>> vtkDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #17
0x00007fffee835c55 in
>>>>>>>>
vtkStreamingDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #18
0x00007fffee816e1a in
>>>>>>>> vtkCompositeDataPipeline::ForwardUpstream(vtkInformation*) ()
from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #19
0x00007fffee81ccb5 in
>>>>>>>> vtkDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #20
0x00007fffee835c55 in
>>>>>>>>
vtkStreamingDemandDrivenPipeline::ProcessRequest(vtkInformation*,
>>>>>>>> vtkInformationVector**, vtkInformationVector*) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #21
0x00007fffee836482 in
>>>>>>>> vtkStreamingDemandDrivenPipeline::Update(int) () from
>>>>>>>> /usr/lib/libvtkCommonExecutionModel.so.1 #22
0x00007ffff1289a76 in
>>>>>>>> vtkAbstractVolumeMapper::GetBounds() () from
>>>>>>>> /usr/lib/libvtkRenderingCore.so.1 #23 0x00007ffff13459f9 in
>>>>>>>> vtkVolume::GetBounds() () from
/usr/lib/libvtkRenderingCore.so.1 #24
>>>>>>>> 0x000000000043f235 in createVolume (image=..., from=0,
>>>>>>>> to=2.7803999378532183, opacityFunction=...,
colorFunction=...) at
>>>>>>>>
>>>>>>>>
>>>>>>>>
/home/estan/Projekt/orexplore/dev/src/insight/src/view/Pipeline.cpp:123 #25
>>>>>>>> 0x00000000004295c4 in CreateVolume::operator() (this=0x829248,
>>>>>>>> image=...) at
>>>>>>>>
/home/estan/Projekt/orexplore/dev/src/insight/src/view/Pipeline.h:45
>>>>>>>> #26
>>>>>>>> 0x000000000042bc7a in
>>>>>>>> QtConcurrent::MappedEachKernel::const_iterator,
>>>>>>>> >::runIteration (this=0x829210, it=...,
>>>>>>>> result=0x7fffbc002da8)
>>>>>>>> at /usr/include/qt/QtConcurrent/qtconcurrentmapkernel.h:176
#27
>>>>>>>> 0x000000000042bd5d in
>>>>>>>> QtConcurrent::MappedEachKernel::const_iterator,
>>>>>>>> >::runIterations (this=0x829210,
>>>>>>>> sequenceBeginIterator=...,
>>>>>>>> begin=1, end=2, results=0x7fffbc002da8) at
>>>>>>>> /usr/include/qt/QtConcurrent/qtconcurrentmapkernel.h:186 #28
>>>>>>>> 0x000000000042c4e1 in
QtConcurrent::IterateKernel::const_iterator,
>>>>>>>> vtkSmartPointer >::forThreadFunction (this=0x829210) at
>>>>>>>> /usr/include/qt/QtConcurrent/qtconcurrentiteratekernel.h:256
#29
>>>>>>>> 0x000000000042bedc in
QtConcurrent::IterateKernel::const_iterator,
>>>>>>>> vtkSmartPointer >::threadFunction (this=0x829210) at
>>>>>>>> /usr/include/qt/QtConcurrent/qtconcurrentiteratekernel.h:218
#30
>>>>>>>> 0x00007ffff7bd5cfd in QtConcurrent::ThreadEngineBase::run()
() from
>>>>>>>> /usr/lib/libQt5Concurrent.so.5 #31 0x00007ffff489a01f in ??
() from
>>>>>>>> /usr/lib/libQt5Core.so.5 #32 0x00007ffff489dd78 in ?? () from
>>>>>>>> /usr/lib/libQt5Core.so.5 #33 0x00007fffeb3f5454 in
start_thread ()
>>>>>>>> from
>>>>>>>> /usr/lib/libpthread.so.0 #34 0x00007fffec5f07df in clone ()
from
>>>>>>>> /usr/lib/libc.so.6 (gdb) Before I start digging into what is
>>>>>>>> happening
>>>>>>>> here
>>>>>>>> I'd just like to ask: Do I have to do something special when
using
>>>>>>>> the
>>>>>>>> HDF5
>>>>>>>> library from two different threads? I'm not reading the same
files
>>>>>>>> or
>>>>>>>> anything, it's simply two completely separate usages of the
library
>>>>>>>> in
>>>>>>>> threads that run in parallel. Does the library have any global
>>>>>>>> structures or
>>>>>>>> something that must be initialized before spawning any
threads that
>>>>>>>> use it?
>>>>>>>> The reason I'm a little worried is that perhaps I've just
been lucky
>>>>>>>> when
>>>>>>>> running under Ubuntu / HDF5 1.8.16. My usage in each thread
>>>>>>>> basically
>>>>>>>> looks
>>>>>>>> like: 1) Create a H5::H5File 2) Open a dataset using
>>>>>>>> file.openDataset
>>>>>>>> 3) Get
>>>>>>>> the dataspace for the dataset and select a hyperslab 4)
Create a
>>>>>>>> memory
>>>>>>>> dataspace 5) Perform a single read(..) operation from the
dataset
>>>>>>>> dataspace
>>>>>>>> to the memory dataspace And it's always different files that
the
>>>>>>>> threads
>>>>>>>> work with. Is there some step 0 I'm missing? Thanks in
advance for
>>>>>>>> any
>>>>>>>> advice. Elvis _______________________________________________
>>>>>>>> Hdf-forum is
>>>>>>>> for HDF software users discussion.
Hdf-forum@lists.hdfgroup.org
>>>>>>>>
>>>>>>>>
>>>>>>>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>>>>>>>> Twitter: https://twitter.com/hdf5
>>>>>>>>
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> Hdf-forum is for HDF software users discussion.
>>>>>>>> Hdf-forum@lists.hdfgroup.org
>>>>>>>>
>>>>>>>>
>>>>>>>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>>>>>>>> Twitter: https://twitter.com/hdf5
>>>>>>
>>>>>> _______________________________________________
>>>>>> Hdf-forum is for HDF software users discussion.
>>>>>> Hdf-forum@lists.hdfgroup.org
>>>>>>
>>>>>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>>>>>> Twitter: https://twitter.com/hdf5
>>>>>
>>>>> _______________________________________________
>>>>> Hdf-forum is for HDF software users discussion.
>>>>> Hdf-forum@lists.hdfgroup.org
>>>>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>>>>> Twitter: https://twitter.com/hdf5
>>>>
>>>> _______________________________________________
>>>> Hdf-forum is for HDF software users discussion.
>>>> Hdf-forum@lists.hdfgroup.org
>>>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
···
Den 8 nov. 2016 4:50 em skrev "Samer Afach" <samer@afach.de>:
On 08.11.2016 4:02 PM, Elvis Stansvik wrote:
2016-11-08 15:46 GMT+01:00 Elvis Stansvik <elvis.stansvik@orexplore.com>:
2016-09-23 7:55 GMT+02:00 Elvis Stansvik <elvis.stansvik@orexplore.com>:
2016-09-22 21:27 GMT+02:00 Werner Benger <werner@cct.lsu.edu>:
> On 22.09.2016 20:34, Elvis Stansvik wrote:
>> 2016-09-22 20:25 GMT+02:00 Werner Benger <werner@cct.lsu.edu>:
>>> On 22.09.2016 20:15, Elvis Stansvik wrote:
>>>> 2016-09-22 19:58 GMT+02:00 Scot Breitenfeld <brtnfld@hdfgroup.org >:
>>>>>> On Sep 22, 2016, at 12:36 PM, Elvis Stansvik >>>> >>>>>> <elvis.stansvik@orexplore.com> wrote:
>>>>>> 2016-09-22 19:23 GMT+02:00 Elvis Stansvik
>>>>>>> 2016-09-22 19:17 GMT+02:00 Dana Robinson <derobins@hdfgroup.org >:
>>>> Twitter: https://twitter.com/hdf5
>>>
>>>
>>> --
>>>
>>>
___________________________________________________________________________
>>> Dr. Werner Benger Visualization Research
>>> Center for Computation & Technology at Louisiana State University
>>> (CCT/LSU)
>>> 2019 Digital Media Center, Baton Rouge, Louisiana 70803
>>> Tel.: +1 225 578 4809 Fax.: +1 225 578-5362
>>>
>>>
>>>
>>> _______________________________________________
>>> Hdf-forum is for HDF software users discussion.
>>> Hdf-forum@lists.hdfgroup.org
>>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>>> Twitter: https://twitter.com/hdf5
>>
>> _______________________________________________
>> Hdf-forum is for HDF software users discussion.
>> Hdf-forum@lists.hdfgroup.org
>>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>> Twitter: https://twitter.com/hdf5
>
>
> --
>
___________________________________________________________________________
> Dr. Werner Benger Visualization Research
> Center for Computation & Technology at Louisiana State University
(CCT/LSU)
> 2019 Digital Media Center, Baton Rouge, Louisiana 70803
> Tel.: +1 225 578 4809 Fax.: +1 225 578-5362
>
>
> _______________________________________________
> Hdf-forum is for HDF software users discussion.
> Hdf-forum@lists.hdfgroup.org
>
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
> Twitter: https://twitter.com/hdf5
_______________________________________________ Hdf-forum is for HDF
software users discussion. Hdf-forum@lists.hdfgroup.org
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
_______________________________________________
Hdf-forum is for HDF software users discussion.
Hdf-forum@lists.hdfgroup.org
http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
Twitter: https://twitter.com/hdf5
