[mpich-discuss] Use of MPI derived data types / MPI file IO

Wei-keng Liao wkliao at ece.northwestern.edu
Sun Nov 18 18:58:37 CST 2012


Hi, John,

You certainly are on the right track to achieve that. Your code is almost
there, only the call to MPI_File_set_view is incorrect. In fact, you don't need it.

Try remove the call to MPI_File_set_view and replace the MPI_File_write_all with:
MPI_File_write_at_all(f, offset, &atoms[0], (int)atoms.size(), mpi_atom_type_resized, &stat);

On the reader side, you need to set the offset based on the new struct. Other than
that, it is the same as the writer case. (no need of MPI_File_set_view either).

As for the portability issue, I would suggest you to use high-level I/O libraries,
such as PnetCDF.

Wei-keng

On Nov 18, 2012, at 12:38 PM, <jgrime at uchicago.edu> <jgrime at uchicago.edu> wrote:

> Hi Wei-keng,
> 
> That's a good point, thanks!
> 
> However, I actually only want to save certain parts of the "atom" structure to file, 
> and saving the whole array as a raw dump could waste a lot of disk space.
> 
> For example, the "atom" structure I used in the example code in reality contains 
> not only an integer and three contiguous doubles, but also at least another two 
> double[3] entries which I may not want to save to disk. As the full data set can 
> be hundreds of millions (or even billions) of "atom" structures, using a derived 
> data type with only a restricted subset of the data in each "atom" structure will 
> produce considerably smaller file sizes!
> 
> There's also the problem of making the resultant file "portable" - raw memory 
> dumps could make life difficult in trying to use output files on machines with 
> different processor architectures. Once I get the derived data types working, I 
> can then switch from the "native" representation to something else ("external32" 
> etc), which should allow me to create portable output files, provided I'm careful 
> with using MPIs file offset routines etc if the file is larger than plain old 32 bit 
> offsets can handle.
> 
> Cheers,
> 
> J.
> 
> ---- Original message ----
>> Date: Sun, 18 Nov 2012 12:27:04 -0600
>> From: discuss-bounces at mpich.org (on behalf of Wei-keng Liao 
> <wkliao at ece.northwestern.edu>)
>> Subject: Re: [mpich-discuss] Use of MPI derived data types / MPI file IO  
>> To: discuss at mpich.org
>> 
>> Hi, John
>> 
>> If your I/O is simply appending one process's data after another and the I/O 
> buffers in memory
>> are all contiguous, then you can simply do the following without defining MPI
>> derived data types or setting the file view.
>> 
>> MPI_File_write_at_all(f, offset, &atoms[0], (int)atoms.size() * sizeof(struct atom), 
> MPI_BYTE, &stat);
>> 
>> Using derived data types is usually when you have noncontiguous buffer in 
> memory or
>> want to access non-contiguous data in files.
>> 
>> 
>> Wei-keng
>> 
>> On Nov 18, 2012, at 11:52 AM, <jgrime at uchicago.edu> 
> <jgrime at uchicago.edu> wrote:
>> 
>>> Hi all,
>>> 
>>> I'm having some problems with using derived data types and MPI parallel IO, 
> and 
>>> was wondering if anyone could help. I tried to search the archives in case 
> this 
>>> was covered earlier, but that just gave me "ht://Dig error" messages.
>>> 
>>> Outline: I have written a C++ program where each MPI rank acts on data 
> stored 
>>> in a local array of structures. The arrays are typically of different lengths on 
> each 
>>> rank. I wish to write and read the contents of these arrays to disk using MPI's 
>>> parallel IO routines. The file format is simply an initial integer which 
> describes 
>>> how many "structures" are in the file, followed by the data which represents 
> the 
>>> "structure information" from all ranks (ie the total data set).
>>> 
>>> So far, I've tried two different approaches: the first consists of each rank 
>>> serialising the contents of the local array of structures into a byte array, 
> which is 
>>> then saved to file "f" using MPI_File_set_view( f, MPI_COMM_WORLD, offset, 
>>> MPI_CHAR, MPI_CHAR, "native", MPI_INFO_NULL ) to skip the initial integer 
>>> "header" and then a call to MPI_File_write_all( f, local_bytearray, 
> local_n_bytes, 
>>> MPI_CHAR, &status ). Here, "offset" is simply the size of an integer (in bytes) 
> + 
>>> the summation of the number of bytes each preceeding rank wishes to write 
> to 
>>> the file (received via an earlier MPI_Allgather call). This seems to work, as 
> when I 
>>> read the file back in on a single MPI rank and deserialise the data into an 
> array of 
>>> structures I get the results I expect.
>>> 
>>> The second approach is to use MPI's derived data types to create MPI 
>>> representations of the structures, and then treat the arrays of structures as 
> MPI 
>>> data types. This allows me to avoid copying the local data into an 
> intermediate 
>>> buffer etc, and seems the more elegant approach. I cannot, however, seem 
> to 
>>> make this approach work.
>>> 
>>> I'm pretty sure the problem lies in my use of the file views, but I'm not sure 
>>> where I'm going wrong. The reading of the integer "header" always works 
> fine, 
>>> but the proceeding data is garbled. I'm using the "native" data representation 
> for 
>>> testing, but will likely change that to something more portable when I get 
> this 
>>> code working.
>>> 
>>> I've included the important excerpts of the test code I'm trying to use below 
>>> (with some printf()s and error handling etc removed to make it a little more 
>>> concise). I have previously tested that std::vector allocates a contiguous flat 
>>> array of the appropriate data type in memory, so passing a pointer/reference 
> to 
>>> the first element in such a data structure behaves the same way as simply 
>>> passing a conventional array of the appropriate data type:
>>> 
>>> struct atom
>>> {
>>> 	int global_id;
>>> 	double xyz[3];
>>> };
>>> 
>>> void write( char * fpath, std::vector<struct atom> &atoms, int rank, int 
> nranks )
>>> {
>>> 	/*
>>> 		Memory layout information for the structure we wish to convert into 
>>> an
>>> 		MPI derived data type.
>>> 	*/
>>> 	std::vector<int> s_blocklengths;
>>> 	std::vector<MPI_Aint> s_displacements;
>>> 	std::vector<MPI_Datatype> s_datatypes;
>>> 	MPI_Aint addr_start, addr;
>>> 	MPI_Datatype mpi_atom_type, mpi_atom_type_resized;
>>> 	int type_size;
>>> 	
>>> 	struct atom a;
>>> 	
>>> 	MPI_File f;
>>> 	MPI_Status stat;
>>> 	MPI_Offset offset;
>>> 	char *datarep = (char *)"native";
>>> 
>>> 	std::vector<int> all_N;
>>> 	int local_N, global_N;
>>> 
>>> 	/*
>>> 		Set up the structure data type: single integer, and 3 double precision 
>>> floats.
>>> 		We use the temporary "a" structure to determine the layout of memory 
>>> inside
>>> 		atom structures.
>>> 	*/
>>> 	MPI_Get_address( &a, &addr_start );
>>> 	
>>> 	s_blocklengths.push_back( 1 );
>>> 	s_datatypes.push_back( MPI_INT );
>>> 	MPI_Get_address( &a.global_id, &addr );
>>> 	s_displacements.push_back( addr - addr_start );
>>> 
>>> 	s_blocklengths.push_back( 3 );
>>> 	s_datatypes.push_back( MPI_DOUBLE );
>>> 	MPI_Get_address( &a.xyz[0], &addr );
>>> 	s_displacements.push_back( addr - addr_start );
>>> 	
>>> 	MPI_Type_create_struct( (int)s_blocklengths.size(), &s_blocklengths[0], 
>>> &s_displacements[0], &s_datatypes[0], &mpi_atom_type );
>>> 	MPI_Type_commit( &mpi_atom_type );
>>> 	
>>> 	/*
>>> 		Take into account any compiler padding in creating an array of 
>>> structures.
>>> 	*/
>>> 	MPI_Type_create_resized( mpi_atom_type, 0, sizeof(struct atom), 
>>> &mpi_atom_type_resized );
>>> 	MPI_Type_commit( &mpi_atom_type_resized );
>>> 		
>>> 	MPI_Type_size( mpi_atom_type_resized, &type_size );
>>> 
>>> 	local_N = (int)atoms.size();
>>> 	all_N.resize( nranks );
>>> 
>>> 	MPI_Allgather( &local_N, 1, MPI_INT, &all_N[0], 1, MPI_INT, 
>>> MPI_COMM_WORLD );
>>> 
>>> 	global_N = 0;
>>> 	for( size_t i=0; i<all_N.size(); i++ ) global_N += all_N[i];
>>> 
>>> 	offset = 0;
>>> 	for( int i=0; i<rank; i++ ) offset += all_N[i];
>>> 
>>> 	offset *= type_size; // convert from structure counts -> bytes into file for 
>>> true structure size
>>> 	offset += sizeof( int ); // skip leading integer (global_N) in file.
>>> 
>>> 	MPI_File_open( MPI_COMM_WORLD, fpath, MPI_MODE_CREATE | 
>>> MPI_MODE_WRONLY, MPI_INFO_NULL, &f );
>>> 	if( rank == 0 )
>>> 	{
>>> 		MPI_File_write( f, &global_N, 1, MPI_INT, &stat );
>>> 	}
>>> 	MPI_File_set_view( f, offset, mpi_atom_type_resized, 
>>> mpi_atom_type_resized, datarep, MPI_INFO_NULL );
>>> 	
>>> 	MPI_File_write_all( f, &atoms[0], (int)atoms.size(), mpi_atom_type_resized, 
>>> &stat );
>>> 	MPI_File_close( &f );
>>> 
>>> 	MPI_Type_free( &mpi_atom_type );
>>> 	MPI_Type_free( &mpi_atom_type_resized );
>>> 
>>> 	return;
>>> }
>>> 
>>> void read( char * fpath, std::vector<struct atom> &atoms )
>>> {
>>> 	std::vector<int> s_blocklengths;
>>> 	std::vector<MPI_Aint> s_displacements;
>>> 	std::vector<MPI_Datatype> s_datatypes;
>>> 	MPI_Datatype mpi_atom_type, mpi_atom_type_resized;
>>> 	
>>> 	struct atom a;
>>> 	MPI_Aint addr_start, addr;
>>> 	
>>> 	MPI_File f;
>>> 	MPI_Status stat;
>>> 	
>>> 	int global_N;
>>> 	char *datarep = (char *)"native";
>>> 
>>> 	int type_size;
>>> 
>>> 	/*
>>> 		Set up the structure data type
>>> 	*/
>>> 	MPI_Get_address( &a, &addr_start );
>>> 	
>>> 	s_blocklengths.push_back( 1 );
>>> 	s_datatypes.push_back( MPI_INT );
>>> 	MPI_Get_address( &a.global_id, &addr );
>>> 	s_displacements.push_back( addr - addr_start );
>>> 
>>> 	s_blocklengths.push_back( 3 );
>>> 	s_datatypes.push_back( MPI_DOUBLE );
>>> 	MPI_Get_address( &a.xyz[0], &addr );
>>> 	s_displacements.push_back( addr - addr_start );
>>> 	
>>> 	MPI_Type_create_struct( (int)s_blocklengths.size(), &s_blocklengths[0], 
>>> &s_displacements[0], &s_datatypes[0], &mpi_atom_type );
>>> 	MPI_Type_commit( &mpi_atom_type );
>>> 	
>>> 	/*
>>> 		Take into account any compiler padding in creating an array of 
>>> structures.
>>> 	*/
>>> 	MPI_Type_create_resized( mpi_atom_type, 0, sizeof(struct atom), 
>>> &mpi_atom_type_resized );
>>> 	MPI_Type_commit( &mpi_atom_type_resized );
>>> 
>>> 	MPI_Type_size( mpi_atom_type_resized, &type_size );
>>> 	
>>> 	MPI_File_open( MPI_COMM_SELF, fpath, MPI_MODE_RDONLY, 
>>> MPI_INFO_NULL, &f );
>>> 
>>> 	MPI_File_read( f, &global_N, 1, MPI_INT, &stat );
>>> 	
>>> 	atoms.clear();
>>> 	atoms.resize( global_N );
>>> 
>>> 	errcode = MPI_File_set_view( f, sizeof(int), mpi_atom_type_resized, 
>>> mpi_atom_type_resized, datarep, MPI_INFO_NULL );
>>> 	errcode = MPI_File_read( f, &atoms[0], global_N, mpi_atom_type_resized, 
>>> &stat );
>>> 	errcode = MPI_File_close( &f );
>>> 
>>> 	MPI_Type_free( &mpi_atom_type );
>>> 	MPI_Type_free( &mpi_atom_type_resized );
>>> 
>>> 	return;
>>> }
>>> 
>>> Calling MPI_Type_get_extent() and MPI_Type_get_true_extent() for both 
>>> mpi_atom_type and mpi_atom_type_resized returns (0,32) bytes in all cases. 
>>> Calling MPI_Type_size() on both derived data types returns 28 bytes.
>>> 
>>> If I call MPI_File_get_type_extent() on both derived data types after opening 
> the 
>>> file, they both resolve to 32 bytes - so I think the problem is in the 
> difference 
>>> between the data representation in memory and on disk. If I explicitly use 32 
>>> bytes in the offset calculation in the write() routine above, it still doesn't 
> work.
>>> 
>>> I'm finding it remarkably difficult to do something very simple using MPI's 
>>> derived data types and the parallel IO, and hence I'm guessing that I have 
>>> fundamentally misunderstood one or more aspects of this. If anyone can 
> help 
>>> clarify where I'm going wrong, that would be much appreciated!
>>> 
>>> Cheers,
>>> 
>>> John.
>>> _______________________________________________
>>> discuss mailing list     discuss at mpich.org
>>> To manage subscription options or unsubscribe:
>>> https://lists.mpich.org/mailman/listinfo/discuss
>> 
>> _______________________________________________
>> discuss mailing list     discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss




More information about the discuss mailing list