Thanks again for taking the time and effort to answer my questions. 'am
going to order that book now.
On Fri, Dec 10, 2010 at 9:13 AM, Jayesh Krishna <jayesh at mcs.anl.gov> wrote:
> Hi,
> All MPI processes are OS processes not threads and each MPI process
> executes the same code. So creating an array in one process should not
> interfere with creating an array in another process.
> You can broadcast the data used to populate the array to other MPI
> processes from the root. The other processes should still have a placeholder
> (an array) to store the data. Since MPI_Bcast() blocks on the communicator
> (as mentioned in my previous email), each MPI process blocks on the call
> till it receives the data.
> Hope this helps.
>
> (PS: You might also want to take a look at books like
> http://www.mcs.anl.gov/research/projects/mpi/usingmpi/ to get an overview
> of MPI)
> Regards,
> Jayesh
> ----- Original Message -----
> From: Prashanth <prashanth.dumpuri at gmail.com>
> To: Jayesh Krishna <jayesh at mcs.anl.gov>
> Cc: mpich-discuss at mcs.anl.gov
> Sent: Thu, 09 Dec 2010 16:43:22 -0600 (CST)
> Subject: Re: [mpich-discuss] MPICH2 hangs during debug
>
> Thanks for all your help. My code works but part of it still has memory
> leaks. Hopefully the solution to that lies in the answer to this
> question(s):
>
> Currently I let all processors create one sequential array and populate
> the same array. Is this thread-safe? Or should I let root node create the
> array and then broadcast it? If I let root node create that array, how do i
> ensure other nodes/processors are not going to jump and execute the
> statements below the array creation?
>
> Thanks again
>
>
> On Thu, Dec 9, 2010 at 4:30 PM, Jayesh Krishna <jayesh at mcs.anl.gov> wrote:
>
> > Hi,
> > Yes, you can execute non-parallel version of class 1 from the root MPI
> > process (rank = 0) and broadcast the result to other MPI processes.
> > Typically you divide work (computation) among MPI processes if the work
> is
> > scalable and the cost of computation of the results using a single worker
> is
> > greater than the (cost of computation of the divided work in a single
> worker
> > + cost of communication of the result among MPI processes). If the work
> is
> > significantly small you might be better off doing it in a single process.
> > Make sure that when you call MPI_Bcast() from all the MPI processes
> > (Correctly speaking, all MPI processes belonging to the same communicator
> > should call MPI_Bcast(). In your case the communicator, MPI_COMM_WORLD,
> > consists of all the MPI processes).
> > From your email I am guessing that you no longer have any problems
> > executing your MPI program with MPICH2 (Am I right ? ).
> >
> > Regards,
> > Jayesh
> > ----- Original Message -----
> > From: Prashanth <prashanth.dumpuri at gmail.com>
> > To: Jayesh Krishna <jayesh at mcs.anl.gov>
> > Cc: mpich-discuss at mcs.anl.gov
> > Sent: Thu, 09 Dec 2010 15:38:55 -0600 (CST)
> > Subject: Re: [mpich-discuss] MPICH2 hangs during debug
> >
> > Jayesh,
> > Just playing catch up in the next 2 lines - I have 3 classes 2 of
> > which have been parallelized using MPI and all 3 are being instantiated
> and
> > executed in my main code. Since one of the classes was not parallelized i
> > had it inside an "if" condition: if rank==zero then execute the
> > non-parallel
> > class1. I then used MPI_BCast to broadcast the results from class1 to all
> > the processors.
> > Turns out this "if" condition caused the child processors (those
> > with rank not equal to zero) to jump directly to the functions below the
> > MPI_BCast command while root node was still stuck at MPI_BCast. This
> caused
> > my code/debugger to hang.
> > It takes me less than 2seconds to execute class1 so I can get
> away
> > with running class1 on all the processors but I want to learn the
> "proper"
> > way to execute a MPI application. Any thoughts?
> > Please let me know if the above email was confusing.
> > Thanks
> > Prashanth
> >
> >
> > On Thu, Dec 9, 2010 at 1:48 PM, Prashanth <prashanth.dumpuri at gmail.com
> > >wrote:
> >
> > > Jayesh,
> > > Thanks for the immediate response (again). You need VTK
> (visualization
> > > toolkit) and PETSc (parallel sparse matrix) libraries to compile my
> code.
> > > Can I include these toolkits in the zip file i'm going to send you?
> > > Thanks
> > > Prashanth
> > >
> > > On Thu, Dec 9, 2010 at 1:44 PM, Jayesh Krishna <jayesh at mcs.anl.gov>
> > wrote:
> > >
> > >> Hi,
> > >> The parameters to MPI_Bcast() look alright. It might be easier for me
> > to
> > >> debug if you can send me the complete code. Can you send a zip of the
> > >> complete code to jayesh at mcs.anl.gov ?
> > >>
> > >> Regards,
> > >> Jayesh
> > >> ----- Original Message -----
> > >> From: Prashanth <prashanth.dumpuri at gmail.com>
> > >> To: Jayesh Krishna <jayesh at mcs.anl.gov>
> > >> Cc: mpich-discuss at mcs.anl.gov
> > >> Sent: Thu, 09 Dec 2010 13:33:08 -0600 (CST)
> > >> Subject: Re: [mpich-discuss] MPICH2 hangs during debug
> > >>
> > >> Jayesh,
> > >> Thanks for the immediate response. The example MPI application ran
> > fine
> > >> in the Debug mode. MPI_BCast did not hang for the example application.
> > To
> > >> answer your question as to why i'm reading the same file from all MPI
> > >> processes, I couldn't figure out how to use MPI_Byte(s) / MPI_Send()
> to
> > >> send
> > >> data from root node to all the child nodes. Once I figure that out
> I'll
> > >> ask
> > >> my code to read the data on the root node and then broadcast it to
> child
> > >> nodes.
> > >> I'm not sure if can attach documents to this mailing list and hence
> > 'am
> > >> just pasting the snippet of my code containing the MPI call. If you
> need
> > >> the
> > >> entire code, please let me know how to send it to you and i'll send it
> > to
> > >> you. The code hangs at the MPI_BCast command in the Debug mode. FYI,
> > I'm
> > >> using another toolkit VTK - visualization toolkit - to format my data.
> > >> Output1 from class1 is in the vtk data format and translates to a
> double
> > >> array in C++.
> > >>
> > >> int main( int argc, char * argv[] )
> > >> {
> > >>
> > >> MPI_Init(&argc,&argv);
> > >> int rank, number_of_processors;
> > >> MPI_Comm_size(MPI_COMM_WORLD,& number_of_processors);
> > >> MPI_Comm_rank(MPI_COMM_WORLD,&rank);
> > >>
> > >> // read inputs on all processors
> > >> // snipped for brevity
> > >>
> > >> // class1 - read the inputs and generate output1: need not be
> > >> parallelized
> > >> vtkSmartPointer<vtkGetSignedClosestPointDistances>
> > >> ComputeSignedClosestPointDistances =
> > >>
> > >>
> > >> vtkSmartPointer<vtkGetSignedClosestPointDistances>::New();
> > >> if ( rank == 0 )
> > >> {
> > >> ComputeSignedClosestPointDistances->SetInput( inputs i just
> > read
> > >> );
> > >> ComputeSignedClosestPointDistances->Update();
> > >> }
> > >> // get output1 from class1
> > >> vtkSmartPointer<vtkDoubleArray> signedclosestpointdistancesbefore =
> > >> vtkSmartPointer<vtkDoubleArray>::New();
> > >>
> > >> signedclosestpointdistancesbefore->DeepCopy(
> > >> ComputeSignedClosestPointDistances->GetSignedClosestPointDistances()
> );
> > >>
> > >> // GetVoidPointer(0) returns the void pointer at 0th element
> > >> // GetNumberOfTuples - size of the array
> > >> MPI_Bcast( signedclosestpointdistancesbefore->GetVoidPointer(0),
> > >> signedclosestpointdistancesbefore->GetNumberOfTuples(), MPI_DOUBLE, 0,
> > >> MPI_COMM_WORLD );
> > >>
> > >> // code snipped for brevity
> > >>
> > >> MPI_Finalize();
> > >>
> > >> return 0;
> > >> }
> > >>
> > >> Thanks again for all your help
> > >> Prashanth
> > >>
> > >>
> > >
> >
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20101210/87a29035/attachment.htm>