[mpich-discuss] MPI One-Sided Communication, and indexed datatype

[Jeff Hammond]

Can't you just put it on the Internet somewhere?  Do you have a
world-readable SVN repo?

Jeff

On Thu, May 31, 2012 at 9:18 PM, Ziaul Haque Olive <mzh.olive at gmail.com> wrote:
> Thanks Jeff,
>
> the code is quite large to post here. It is from graph500 benchmark's mpi
> implementation. multiple files are required to build this code. I can send
> the whole directory for debugging. what do you say?
>
> Thanks,
> Ziaul.
>
>
> On Thu, May 31, 2012 at 8:51 PM, Jeff Hammond <jhammond at alcf.anl.gov> wrote:
>>
>> This is an interesting question.
>>
>> "Is it okay that I am freeing memory before the closing fence?"
>>
>> With respect to the memory associated with the MPI_Accumulate itself,
>> the answer is clearly "no".
>>
>> MPI-2.2 11.3:
>>
>> "The local communication buffer of an RMA call should not be updated,
>> and the local communication buffer of a get call should not be
>> accessed after the RMA call, until the subsequent synchronization call
>> completes."
>>
>> Freeing memory is obviously a modification.
>>
>> However, it seems you are freeing the MPI_Datatype used in this
>> operation and the input arrays associated with it.
>>
>> From MPI-2.2 4.1.9:
>>
>> "Marks the datatype object associated with datatype for deallocation
>> and sets datatype to MPI_DATATYPE_NULL. Any communication that is
>> currently using this datatype will complete normally. Freeing a
>> datatype does not affect any other datatype that was built from the
>> freed datatype. The system behaves as if input datatype arguments to
>> derived datatype constructors are passed by value."
>>
>> It would seem that your usage is correct.
>>
>> I cannot debug your code because it is incomplete and has a few bugs
>> according to the C standard.  Can you post the code that actually
>> results in the error you're seeing?
>>
>> There may be other issues with your code that others may point out.
>> My email is not intended to capture all possible sources of error.
>>
>> Best,
>>
>> Jeff
>>
>>
>> On Thu, May 31, 2012 at 6:23 PM, Ziaul Haque Olive <mzh.olive at gmail.com>
>> wrote:
>> > Hello,
>> >
>> > I am not sure, if my code is correct according to MPICH2(v1.4.1p1.). the
>> > code is given as follows, I am doing MPI one-sided communication inside
>> > a
>> > function - data_transfer. this function is being called inside a fence
>> > epoch. inside data_transfer, I am allocating memory for non-contiguous
>> > data,
>> > creating derived data type, using this datatype in MPI_Accumulate, and
>> > after
>> > calling MPI_Accumulate, freeing the indexed data type and also freeing
>> > the
>> > memory containing indices for indexed data type. is it okay that I am
>> > freeing memory before the closing fence.
>> >
>> > void data_transfer(void *data, int *sources_disp, int *targets_disp, int
>> > *target, int size, int *blength, int func, MPI_Op op, MPI_Win win,
>> > MPI_Datatype dtype){
>> >
>> > int i,j, index;
>> > int tmp_target;
>> > int *flag;
>> > int *source_disp;
>> > int *target_disp;
>> > MPI_Datatype source_type, target_type;
>> > MPI_Alloc_mem( size*sizeof(int), MPI_INFO_NULL, &source_disp);
>> > MPI_Alloc_mem( size*sizeof(int), MPI_INFO_NULL, &target_disp);
>> > MPI_Alloc_mem( size*sizeof(int), MPI_INFO_NULL, &flag );
>> >
>> > memset(flag, 0, size*sizeof(int));
>> >
>> > for(i=0;i<size;i++){
>> > if(flag[i]==0){
>> > tmp_target = target[i];
>> > index = 0;
>> > for(j=i; j<size; j++){
>> > if(flag[j]==0 && tmp_target == target[j] ){
>> > source_disp[index] = sources_disp[j];
>> > target_disp[index] = targets_disp[j];
>> > //printf("src, target disp %d  %d\n", j, disp[j]);
>> > index++;
>> > flag[j] = 1;
>> > }
>> > }
>> >
>> > MPI_Type_indexed(index, blength , source_disp, dtype, &source_type);
>> > MPI_Type_commit(&source_type);
>> > MPI_Type_indexed(index, blength , target_disp, dtype, &target_type);
>> > MPI_Type_commit(&target_type);
>> > MPI_Accumulate( data, 1, source_type, tmp_target, 0, 1, target_type ,
>> > op,
>> > win);
>> >
>> > MPI_Type_free(&source_type);
>> > MPI_Type_free(&target_type);
>> > }
>> > }
>> > MPI_Free_mem(source_disp);
>> > MPI_Free_mem(target_disp);
>> > MPI_Free_mem(flag);
>> >
>> > }
>> >
>> > void main(){
>> > int i;
>> > while(i<N){
>> > MPI_Win_fence(MPI_MODE_NOPRECEDE, queue2_win);
>> >
>> > data_transfer();
>> >
>> > MPI_Win_fence(MPI_MODE_NOSUCCEED, queue2_win);
>> > }
>> > }
>> >
>> > Thanks
>> > Ziaul
>> >
>> > _______________________________________________
>> > mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>> > To manage subscription options or unsubscribe:
>> > https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>> >
>>
>>
>>
>> --
>> Jeff Hammond
>> Argonne Leadership Computing Facility
>> University of Chicago Computation Institute
>> jhammond at alcf.anl.gov / (630) 252-5381
>> http://www.linkedin.com/in/jeffhammond
>> https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond
>> _______________________________________________
>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>> To manage subscription options or unsubscribe:
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>
>
>
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond