MPI is only used if it's available. It tries to see if there is a
MPIController available. If not, then the number of ranks is 1, which
means that the parallel class calls the super class's requestdata (super
class is the serial wind blade reader).
However, the parallel class overrides certain functions to make use of MPI.
So even the parallel instantiation will call the superclass's requestdata,
the overriden functions that are called from that point on assume that MPI
is available...this is a problem when the MPI controller is not available.
On another note, it appears that the rectilinear grid that I am reading
(VTK file, structured points I think) is read by the vtkFileSeriesReader,
which makes use of the vtkPDataSetReader class. Strange that if I use
vtkPDataSetReader myself in C++ code that uses VTK and not paraview,
bounds/cell information is not generated. So I think that ParaView does
something else besides using that reader...hmmm.
On Mon, Aug 27, 2012 at 7:32 PM, Andy Bauer <andy.ba...@kitware.com> wrote:
>
>
> On Mon, Aug 27, 2012 at 6:53 PM, Sohail Shafii <sssha...@ucdavis.edu>wrote:
>
>> I see...so if MPI is running then it will create a parallel version of
>> the class. Otherwise, it will create the normal reader.
>>
>
> Technically, if ParaView is built with MPI enabled it will create the
> parallel version of the class every time whether or not the client, server,
> or python scripts are actually run with mpiexec, mpirun, or whatever is
> used to start the executable. So you can't rely on MPI being initialized
> and should not initialize it yourself in your class as that can cause
> problems as well. I think that's what you meant but wanted to be as clear
> as possible for the implications for the classes.
>
>
>>
>> As far as the check goes for a multiprocess controller; while it does
>> call the parent request data in the serial case, the inherited parallel
>> class overrides some of the methods. Which is a problem because it refers
>> to its own file pointer in those methods (it uses MPI_File instead of a
>> FILE*). Hmm.
>>
>> Sohail
>>
>>
>> On Thu, Aug 23, 2012 at 3:20 PM, Andy Bauer <andy.ba...@kitware.com>wrote:
>>
>>> Hi Sohail,
>>>
>>> Sorry for the slow reply but I wanted to make sure how things should
>>> work first and that took a bit to figure out. Anyways, this is closer to
>>> what's needed but the way it is supposed to work in VTK is that we use
>>> vtkObjectFactory to determine which one should be created when
>>> vtkWindBladeReader::New() is called. This is done at build time based on
>>> whether or not VTK is built with MPI. So if VTK and ParaView are built in
>>> parallel then there is code added that creates a vtkPWindBladeReader when
>>> vtkWindBladeReader::New() is called. This is regardless of whether or not
>>> the executable is run with mpi initialized or not. Because of this, the
>>> parallel version of the reader is responsible for checking whether mpi is
>>> initialized, ideally by doing
>>> vtkMultiProcessController::GetGlobalController()->IsA("vtkMPIController")
>>> is true, and if it isn't it should just call the parent class's methods for
>>> reading in the files. You may want to look at vtkPNrrdReader and the
>>> CMakeLists.txt file in that directory to see how they did things. Note
>>> though that if it's a single process running that it doesn't do any MPI
>>> calls even in the parallel version so your code will be slightly different
>>> than theirs.
>>>
>>> As for other parts of the email below:
>>>
>>> 1) yes, please keep cc'ing the list as there are some details in here
>>> that will be helpful to other people trying to implement their own parallel
>>> readers or filters
>>>
>>> 2) I'm not sure what's going on with the gradient filter but am guessing
>>> that it has to do with the multiblock data set and that there probably
>>> isn't that same array in each block. But that's just a guess. Did you try
>>> extracting the "air" block and do the vorticity and/or q criterion
>>> computation on that?
>>>
>>> 3) vtkSMReaderFactory may be the class you're looking for to determine
>>> which reader is getting used. ParaViewReaders.xml has a list of readers
>>> that it will try to read in a file.
>>>
>>> Andy
>>>
>>>
>>> On Tue, Aug 21, 2012 at 7:11 PM, Sohail Shafii <sssha...@ucdavis.edu>wrote:
>>>
>>>> <sent via another e-mail address; yahoo prevents me from uploading
>>>> stuff because it's stupid>
>>>>
>>>> Here's the first version. Do you want me to keep ccing the list? Not
>>>> sure if that's necessary at this point because this is related to the
>>>> development of ParaView and not help.
>>>>
>>>> I've run it with an older version of the master, and it seems to read
>>>> in a test .wind file just fine. For now I had to store the
>>>> vtkPWindBladeReader* files into VTK/Parallel/MPI since I don't have
>>>> VTK/IO/MPIParallel module yet. Now when I run the gradient filter after
>>>> the field, the gradient filter thinks that the input array does not have
>>>> the appropriate number of components for Q criterion or vorticity...but it
>>>> should (UVW does have three components per tuple). Might be my outdated
>>>> copy of the git repo. Haven't debugged that gradient code yet.
>>>>
>>>> *I was wondering if this is the type of code that you want? Or should
>>>> the classes be designed differently?* Also, I have only run it
>>>> parallel because if I try to instantiate it inside of the python shell in
>>>> serial it tries to run RequestInformation (which assumes that the file path
>>>> exists but does not). It's a work-in-progress but it's a start.
>>>>
>>>> A random question: if ParaView reads a file, where would I put a
>>>> breakpoint that tells me which reader it ends up picking?? I have some
>>>> separate VTK code that reads in a rectilinear grid, but does not generate
>>>> cell/bounds information. ParaView does generate information about the
>>>> bounds though.
>>>>
>>>> Sohail
>>>>
>>>> ------------------------------
>>>> *From:* David E DeMarle <dave.dema...@kitware.com>
>>>> *To:* Sohail Shafii <sohailsha...@yahoo.com>
>>>> *Cc:* Andy Bauer <andy.ba...@kitware.com>; "migic...@gmail.com" <
>>>> migic...@gmail.com>; "paraview@paraview.org" <paraview@paraview.org>;
>>>> "Woodring, Jonathan L" <woodr...@lanl.gov>
>>>> *Sent:* Friday, July 20, 2012 10:02 AM
>>>> *Subject:* Re: [Paraview] ParaView thinks that it does not have a wind
>>>> blade reader
>>>>
>>>> Thanks.
>>>>
>>>> In today's git master it now sits in VTK/IO/Parallel/vtkWindbladeReader.
>>>>
>>>> The serial version should go in VTK/IO/Geometry.
>>>>
>>>> The MPI version should go in a new module VTK/IO/MPIParallel.
>>>> (VTK/IO/Parallel depends only on vtkSocketController now and should
>>>> not depend on vtkMPIController).
>>>>
>>>> Finally when that is done we can remove the flag in paraview's xml
>>>> that prevents it from being used in serial.
>>>>
>>>> David E DeMarle
>>>> Kitware, Inc.
>>>> R&D Engineer
>>>> 21 Corporate Drive
>>>> Clifton Park, NY 12065-8662
>>>> Phone: 518-881-4909
>>>>
>>>>
>>>> On Thu, Jul 19, 2012 at 7:08 PM, Sohail Shafii <sohailsha...@yahoo.com>
>>>> wrote:
>>>> > Well it seems like #1 and #2 can be solved by writing a new parallel
>>>> wind
>>>> > blade reader class. I had similar experience by creating distributed
>>>> > versions of class that I created in other vtk directories. The normal
>>>> wind
>>>> > blade reader would have no parallel calls, which would address #1.
>>>> >
>>>> > What I find interesting is that in your directories, the wind blade
>>>> reader
>>>> > is located in VTK/IO/Parallel. For the ParaView snapshot that I
>>>> downloaded,
>>>> > as well as the 3.14 source, it's still in VTK/Parallel. No Parallel
>>>> folder
>>>> > in VTK/IO. Is this a recent change?
>>>> >
>>>> > So maybe the ideal folder layout would be:
>>>> >
>>>> > VTK/IO/WindBladeReader.*
>>>> >
>>>> > VTK/IO/Parallel/DistributedWindBladeReader.*
>>>> >
>>>> > I can make changes to the WindBladeReader to remove the MPI-IO calls
>>>> and
>>>> > make a distributed wind blade reader class in case you want to add
>>>> them to
>>>> > the repository. The one template I have for the parallel/serial split
>>>> is the
>>>> > vtkStreamTracer and its distributed counterparts so I can use those
>>>> as a
>>>> > guide. If you have better recommendations please let me know.
>>>> >
>>>> > Sohail
>>>> >
>>>> > ________________________________
>>>> > From: David E DeMarle <dave.dema...@kitware.com>
>>>> > To: Sohail Shafii <sohailsha...@yahoo.com>
>>>> > Cc: Andy Bauer <andy.ba...@kitware.com>; "migic...@gmail.com"
>>>> > <migic...@gmail.com>; "paraview@paraview.org" <paraview@paraview.org
>>>> >;
>>>> > "Woodring, Jonathan L" <woodr...@lanl.gov>
>>>> > Sent: Thursday, July 19, 2012 3:20 PM
>>>> >
>>>> > Subject: Re: [Paraview] ParaView thinks that it does not have a wind
>>>> blade
>>>> > reader
>>>> >
>>>> > Two issues need to be fixed.
>>>> > 1) that behavior in 3.14 is there because of this fix which we had to
>>>> > rush in just before the 3.14 release.
>>>> >
>>>> > Committer: Utkarsh Ayachit <utkarsh.ayac...@kitware.com> 2012-02-08
>>>> > 17:12:18
>>>> > Parent: e401d569f6e54c49c3497852463f6b13a01da8b6 (Update version after
>>>> > 3.14.0-RC1 tag.)
>>>> > Child: 99fdb1aeb5997c29b43393f45a478c4b53a23109 (Merge topic
>>>> > 'fix_windblade_segfault')
>>>> > Branches: many (31)
>>>> > Follows: v3.14.0-RC1
>>>> > Precedes: v3.14.0-RC2
>>>> >
>>>> > Fix segfault with ParaView built with MPI when opening *.vtr files.
>>>> >
>>>> > The problem was that vtkWindBlade reader was getting created to
>>>> test
>>>> > "CanReadFile" and that would cause segfault unless MPI was
>>>> initialized.
>>>> >
>>>> > Solved the problem by added support to vtkSMReaderFactory to
>>>> respect the
>>>> > ProcessSupport flag on SourceProxy and change windblade reader XML
>>>> > to say that
>>>> > it only works in parallel.
>>>> >
>>>> > In other words, Windblade reader's Constructor should not take it upon
>>>> > itself to initialize MPI.
>>>> >
>>>> > Once that is fixed, we get to issue 2 before paraview 4.0
>>>> >
>>>> > 2) VTK's new modularization has the rule that no module can behave
>>>> > differently with parallel (or most other cmake config options) turned
>>>> > configured on or off. This is so because it makes it very hard to
>>>> > package VTK when the libs behave differently in different situations.
>>>> >
>>>> > Ie, we can't have any more
>>>> > #ifdef VTK_USE_MPI
>>>> > do something
>>>> > #else
>>>> > do something different
>>>> > #endif
>>>> > in VTK classes.
>>>> >
>>>> > Now if we want a class to behave differently when MPI is turned on or
>>>> > off we have to make a serial only version in a non MPI dependent
>>>> > module and a subclass of that in an MPI dependent module.
>>>> >
>>>> > David E DeMarle
>>>> > Kitware, Inc.
>>>> > R&D Engineer
>>>> > 21 Corporate Drive
>>>> > Clifton Park, NY 12065-8662
>>>> > Phone: 518-881-4909
>>>> >
>>>> >
>>>> > On Thu, Jul 19, 2012 at 3:52 PM, Sohail Shafii <
>>>> sohailsha...@yahoo.com>
>>>> > wrote:
>>>> >> When running serially, vtkSMReaderFactory::CanReadFile sees that the
>>>> >> session
>>>> >> has a single process. As such, on line 660 it realizes that the
>>>> source
>>>> >> proxy's process support (in this case the source is a
>>>> vtkWindBladeReader
>>>> >> proxy) is multiple processes, which causes a return of "false" from
>>>> that
>>>> >> function. This makes ParaView think that the wind blade reader
>>>> can't read
>>>> >> a
>>>> >> .wind file in serial. According to Andy, the reader works find in
>>>> >> parallel.
>>>> >>
>>>> >> I assume that a fix for this is pretty trivial? It never did this
>>>> before
>>>> >> in
>>>> >> previous versions of ParaView and I think that the wind blade reader
>>>> code
>>>> >> has remained mostly the same. Andy mentioned some details in person
>>>> when
>>>> >> he
>>>> >> was talking with Jon but at that time I didn't know as much as I do
>>>> now.
>>>> >> There is a mention below about enabling the filter in serial and
>>>> parallel
>>>> >> --
>>>> >> is this done in a xml file?
>>>> >>
>>>> >> Sohail
>>>> >>
>>>> >> ________________________________
>>>> >> From: David E DeMarle <dave.dema...@kitware.com>
>>>> >> To: Sohail Shafii <sohailsha...@yahoo.com>
>>>> >> Cc: Andy Bauer <andy.ba...@kitware.com>; "migic...@gmail.com"
>>>> >> <migic...@gmail.com>; "paraview@paraview.org" <paraview@paraview.org
>>>> >;
>>>> >> "Woodring, Jonathan L" <woodr...@lanl.gov>
>>>> >> Sent: Tuesday, July 17, 2012 7:00 PM
>>>> >>
>>>> >> Subject: Re: [Paraview] ParaView thinks that it does not have a wind
>>>> blade
>>>> >> reader
>>>> >>
>>>> >> Each file that has an association with that file extension, made via
>>>> the
>>>> >> XML, has its CanReadFile method called. Put a breakpoint there to
>>>> >> determine
>>>> >> if ParaView even knows about that reader.
>>>> >>
>>>> >> On Tuesday, July 17, 2012, Sohail Shafii wrote:
>>>> >>
>>>> >> Sorry, but I meant to ask a different question. Here's what Andy
>>>> said
>>>> >> before, which relates to my concern:
>>>> >> ">
>>>> >>> I went through the code and it jogged my memory about previous
>>>> problems
>>>> >>> with
>>>> >>> it. It had an explicit MPI call in the constructor that would cause
>>>> >>> problems
>>>> >>> when paraview wasn't getting built with MPI. Any time a user tried
>>>> to
>>>> >>> open
>>>> >>> an unknown file it would try to create a windblade reader to see if
>>>> the
>>>> >>> reader could handle the file and this would cause a crash. If you
>>>> launch
>>>> >>> a
>>>> >>> separate pvserver the wind blade reader will be available in the
>>>> gui."
>>>> >>
>>>> >> What I meant was; which ParaView class "creates" or "tests" a reader
>>>> to
>>>> >> see
>>>> >> if it can handle a file? The test code you specified has a
>>>> predefined data
>>>> >> set that is loaded in, and I think is more of an example to your
>>>> average
>>>> >> user. Correct me if I'm wrong.
>>>> >>
>>>> >> So basically ParaView will recognize the file extension, tries to
>>>> load it,
>>>> >> crashes somewhere, then asks you specify the data format from a
>>>> dialog.
>>>> >> I'm
>>>> >> trying to figure out where it crashes so that I can see how that
>>>> relates
>>>> >> to
>>>> >> the wind blade reader class and any MPI calls that are in there.
>>>> >>
>>>> >> Sohail
>>>> >>
>>>> >> Subject: Re: [Paraview] ParaView thinks that it does not have a wind
>>>> blade
>>>> >> reader
>>>> >>
>>>> >> kargad:ParaView demarle$ find . -name "*WindBlade*" -print
>>>> >> ---> ./VTK/IO/Parallel/Testing/Cxx/TestWindBladeReader.cxx
>>>> >> ./VTK/IO/Parallel/vtkWindBladeReader.cxx
>>>> >> ./VTK/IO/Parallel/vtkWindBladeReader.h
>>>> >>
>>>> >> David E DeMarle
>>>> >> Kitware, Inc.
>>>> >> R&D Engineer
>>>> >> 21 Corporate Drive
>>>> >> Clifton Park, NY 12065-8662
>>>> >> Phone: 518-881-4909
>>>> >>
>>>> >>
>>>> >> On Tue, Jul 17, 2012 at 2:09 PM, Sohail Shafii <
>>>> sohailsha...@yahoo.com>
>>>> >> wrote:
>>>> >>> What is the class that is responsible for testing to reader to see
>>>> if it
>>>> >>> can
>>>> >>> open the .wind file?
>>>> >>>
>>>> >>> Sohail
>>>> >>>
>>>> >>> ________________________________
>>>> >>> From: Andy Bauer <andy.ba...@kitware.com>
>>>> >>> To: Sohail Shafii <sohailsha...@yahoo.com>
>>>> >>> Cc: "Woodring, Jonathan L" <woodr...@lanl.gov>; Sebastien Jourdain
>>>> >>> <sebastien.jourd...@kitware.com>; "migic...@gmail.com"
>>>> >>> <migic...@gmail.com>;
>>>> >>> "paraview@paraview.org" <paraview@paraview.org>
>>>> >>> Sent: Wednesday, July 11, 2012 10:53 AM
>>>> >>>
>>>> >>> Subject: Re: [Paraview] ParaView thinks that it does not have a wind
>>>> >>> blade
>>>> >>> reader
>>>> >>>
>>>> >>> I was reluctant to enable the reader in serial (i.e. not using MPI
>>>> IO
>>>> >>> which
>>>> >>> I think is getting done by default with the parallel build) because
>>>> of
>>>> >>> the
>>>> >>> following warnings that I got:
>>>> >>>
>>>> >>> Warning: In /home/acbauer/CODE/ParaView/
>>>> >>> ParaView/VTK/IO/Parallel/vtkWindBladeReader.cxx, line 942
>>>> >>> vtkWindBladeReader (0x3a0c800): WindBladeReader error reading file:
>>>> >>> /home/acbauer/DATA/VTKLargeData/Data/WindBladeReader/test1_topo.wind
>>>> >>> Premature EOF while reading block of data. Expected 896000 but got 0
>>>> >>>
>>>> >>> I just assumed that it was getting bad results so I didn't check the
>>>> >>> serial
>>>> >>> output with the parallel output. In either case the code should be
>>>> >>> changed
>>>> >>> before enabling it in serial so that it reads the files correctly
>>>> and
>>>> >>> doesn't produce any warning messages.
>>>> >>>
>>>> >>> Andy
>>>> >>>
>>>> >>> On Wed, Jul 11, 2012 at 12:04 PM, Sohail Shafii <
>>>> sohailsha...@yahoo.com>
>>>> >>> wrote:
>>>> >>>
>>>> >>> I'm not aware of any option to build/link against mpi-io. This is
>>>> just a
>>>> >>> standard build with mpi support that suffers from this issue.
>>>> >>>
>>>> >>> The ParaView website's builds also have the same bug.
>>>> >>>
>>>> >>> ________________________________
>>>> >>> From: "Woodring, Jonathan L" <woodr...@lanl.gov>
>>>> >>> To: Sohail Shafii <sohailsha...@yahoo.com>; Andy Bauer
>>>> >>> <andy.ba...@kitware.com>; Sebastien Jourdain
>>>> >>> <sebastien.jourd...@kitware.com>
>>>> >>> Cc: "migic...@gmail.com" <migic...@gmail.com>; "
>>>> paraview@paraview.org"
>>>> >>> <paraview@paraview.org>
>>>> >>> Sent: Tuesday, July 10, 2012 10:06 AM
>>>> >>>
>>>> >>> Subject: Re: [Paraview] ParaView thinks that it does not have a wind
>>>> >>> blade
>>>> >>> reader
>>>> >>>
>>>> >>> I added the MPI calls to be in sync with what was being added in the
>>>> >>> simulation.
>>>> >>>
>>>> >>> In the SuperComputing demo last year, we updated both the HiGrad
>>>> FireTec
>>>> >>> to
>>>> >>> use MPI-IO and likewise ParaView to do the same thing.
>>>> >>>
>>>> >>> It w
>>>> >>
>>>> >>
>>>> >>
>>>> >> --
>>>> >> David E DeMarle
>>>> >> Kitware, Inc.
>>>> >> R&D Engineer
>>>> >> 21 Corporate Drive
>>>> >> Clifton Park, NY 12065-8662
>>>> >> Phone: 518-881-4909
>>>> >>
>>>> >>
>>>> >
>>>> >
>>>>
>>>>
>>>>
>>>
>>
>
_______________________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
Please keep messages on-topic and check the ParaView Wiki at:
http://paraview.org/Wiki/ParaView
Follow this link to subscribe/unsubscribe:
http://www.paraview.org/mailman/listinfo/paraview
