Jeff Squyres wrote:
> On Oct 8, 2004, at 10:07 AM, Sriram Rallabhandi wrote:
>
>> // These define the maximum sizes of the arrays in the structure
>> #define max1 1000
>> #define max2 80
>> #define max3 10
>> #define max4 50
>>
>> typedef struct
>> {
>> int g1[max1], // g1 array not to exceed max1 items. Actual size
>> is entered by user : psizec
>> rank,
>> flag;
>> float x1[max2], // Actual size : psizev
>> x2[max2]; // Actual size psizev2
>> float f[max3], // Actual size: psizef
>> c[max4], // Actual size : psizeco
>> clen,
>> error;
>> } individual;
>>
>> MPI_Datatype Individ,oldtypes[2];
>> int blockcounts[2];
>> MPI_Aint offsets[2],extent;
>>
>> offsets[0]=0;
>> oldtypes[0] = MPI_INT;
>> blockcounts[0] = 2+psizec;
>> MPI_Type_extent(MPI_INT,&extent);
>> offsets[1] = (2+psizec)*extent;
>> oldtypes[1] = MPI_FLOAT;
>> blockcounts[1] = psizev+psizev2+psizef+psizeco+2;
>
>
> A minor quibble: you can't know for sure that the compiler will lay out
> the individual struct in memory with all the ints (contiguously) first
> and all the floats (contiguously) second. It *probably* will, but
> there is no guarantee of that. If it doesn't, your datatype will not
> match it.
I didn't look too close when this was first posted. I am lazy and
figured someone else would.
In additin to what Jeff said, it looks like you are only accounting for
the number of locations used in each field rather than the number
allocated. That is, the g1 field of the struct actually allocates space
for max1 integers, but you are only accounting for psizec when you setup
your datatype. Same for x1, x2, f, and c.
Hope this helps.
Dave.
>
>> MPI_Type_struct(2,blockcounts,offsets,oldtypes,&Individ);
>> MPI_Type_commit(&Individ);
>
>
> Have you tried instantiating an individual, filling it with data, and
> doing a simple send/recv with it to test whether your datatype works?
>
>> Further, I have to do an MPI_Alltoall operation exchanging
>> "individuals" between nodes. I do the following:
>>
>> MPI_Comm_rank(MPI_COMM_WORLD,&rank);
>> MPI_Comm_size(MPI_COMM_WORLD,&psize);
>>
>> individual *sendbuf, *recvbuf;
>> int *sendcnt, *recvcnt, recv;
>> int *sdisp, *rdisp;
>>
>> // Stop all processes at this point so that all nodes proceed from
>> here simultaneously
>> MPI_Barrier(MPI_COMM_WORLD);
>>
>> // ppsize is the number of individuals with each node.
>> ktmp1 = (floor) (ppsize/psize);
>> kslct = (psize-1)*ktmp1;
>>
>> indx1 = (int *) calloc(ppsize,sizeof(int));
>>
>> // Randomly shuffle the location of individuals and place it in send
>> buffer
>> // Have to allocate memory for send buffer and Recv buffer
>> shuffle_index(&indx1,ppsize);
>>
>> // Create a send buffer of a group of individuals
>> sendbuf = (individual *) malloc(ppsize*sizeof(individual));
>>
>> // Put data into the sendbuf array
>> for (pp=0;pp<ppsize;pp++) {
>> sendbuf[pp] = oldpop.ind[indx1[pp]];
>> }
>>
>> recv = ppsize-((psize-1)*ktmp1);
>> sendcnt = (int *) calloc(psize,sizeof(int));
>> recvcnt = (int *) calloc(psize,sizeof(int));
>>
>> for (pp=0;pp<psize;pp++) {
>> if (rank==pp) {
>> sendcnt[pp] = recv;
>> }
>> else {
>> sendcnt[pp] = ktmp1;
>> }
>> }
>>
>> // I think this call would tell the nodes how much info is coming
>> //So, every node gets recvnnt
>> MPI_Alltoall(sendcnt,1,MPI_INT,recvcnt,1,MPI_INT,MPI_COMM_WORLD);
>>
>> sdisp[0]=0;
>> for (pp=1;pp<psize;pp++) {
>> sdisp[pp] = sendcnt[pp-1]+sdisp[pp-1];
>> }
>> rdisp[0]=0;
>> for (pp=1;pp<psize;pp++) {
>> rdisp[pp] = recvcnt[pp-1]+rdisp[pp-1];
>> }
>>
>> recvbuf = (individual *) malloc(ppsize*sizeof(individual));
>> // Each node scatters some of its individuals to other nodes
>> including itself
>>
>> MPI_Alltoallv(sendbuf,sendcnt,sdisp,Individ,recvbuf,recvcnt,rdisp,Indiv
>> id,MPI_COMM_WORLD);
>>
>> // Now use recvbuf
>> for (pp=0;pp<popsize;pp++) {
>> oldpop.ind[pp] = recvbuf[pp];
>> }
>>
>> Is there anything wrong with the above code?
>
>
> A cursory look didn't reveal any problems (admittedly I didn't look
> closely). Can you be specific about why you are asking? For example,
> are you seeing problems or running into errors?
>
--
Dr. David Cronk, Ph.D. phone: (865) 974-3735
Research Leader fax: (865) 974-8296
Innovative Computing Lab
http://www.cs.utk.edu/~cronk
University of Tennessee, Knoxville
|
