October 2014 - discuss - mailman3.cels.anl.gov

Re: [mpich-discuss] where mpi waitsome only returns one request each time
by Junchao Zhang 19 Oct '14

19 Oct '14

Try to run on another machine, you may get different results. My configure is $ mpichversion MPICH Version: 3.1.3 MPICH Release date: unreleased development copy MPICH Device: ch3:nemesis MPICH configure: --enable-romio --enable-nemesis-dbg-localoddeven --enable-g=all --enable-fast=O0 --enable-fortran=f77,fc CC=gcc CXX=g++ F77=gfortran FC=gfortran MPICH CC: gcc -g -O0 MPICH CXX: g++ -g -O0 MPICH F77: gfortran -g -O0 MPICH FC: gfortran -g -O0 --Junchao Zhang On Sun, Oct 19, 2014 at 9:23 PM, myself <chcdlf(a)126.com> wrote: > > I run several times of the program, the results are as the same as before. Could > you tell me your MPICH version and compiling parameters? The version of > my MPICH is 3.1.3, I compile it as > > $ mpirun --version > HYDRA build details: > Version: 3.1.3 > Release Date: Wed Oct 8 09:37:19 CDT 2014 > CC: gcc -fPIC > CXX: g++ > F77: gfortran > F90: gfortran > Configure options: '--disable-option-checking' > '--prefix=/usr/local/mpich3' '--with-device=ch3:nemesis' 'CFLAGS=-fPIC -O2' > '--disable-fortran' '--cache-file=/dev/null' '--srcdir=.' 'CC=gcc' > 'LDFLAGS= ' 'LIBS=-lpthread ' 'CPPFLAGS= > -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/mpl/include > -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/mpl/include > -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/openpa/src > -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/openpa/src > -D_REENTRANT > -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/mpi/romio/include' > Process Manager: pmi > Launchers available: ssh rsh fork slurm ll lsf sge > manual persist > Topology libraries available: hwloc > Resource management kernels available: user slurm ll lsf sge pbs > cobalt > Checkpointing libraries available: > Demux engines available: poll select > > the whole program is > ========= > > #include "mpi.h" > #include <stdio.h> > #include <stdlib.h> > #include <sys/time.h> > > typedef unsigned char byte; > > int main(int argc, char *argv[]){ > int rank,size; > byte buf[10][10]; > MPI_Request req[10]; > MPI_Status stat[10]; > int outcount; > int array_of_indices[10]; > int i; > int complete = 0; > > MPI_Init(&argc, &argv); > MPI_Comm_size(MPI_COMM_WORLD, &size); > MPI_Comm_rank(MPI_COMM_WORLD, &rank); > > if(rank==0) { > for(i =0; i< 10;i++){ > MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, > &req[i]); > } > printf("post send\n"); > }else { > for(i=0; i< 10;i++){ > MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, > &req[i]); > } > printf("post recv\n"); > } > > for (;;){ > MPI_Waitsome(10, req, &outcount, array_of_indices, stat); > complete += outcount; > printf("rank %d %d index %d\n", rank, outcount, > array_of_indices[0]); > if(complete == 10) > break; > } > printf("rank %d complete\n", rank); > MPI_Finalize(); > } > ========= > > I compile it as > $ mpicc -o tests testsome.c > > and run as > $ mpirun -n 2 ./tests > > > At 2014-10-20 09:02:39, "Junchao Zhang" <jczhang(a)mcs.anl.gov> wrote: > > I think it is random. For example, I ran your code and its output was: > > post recv > post send > rank 0 10 index 0 > rank 0 complete > rank 1 1 index 0 > rank 1 9 index 1 > rank 1 complete > > --Junchao Zhang > > On Sun, Oct 19, 2014 at 9:48 AM, myself <chcdlf(a)126.com> wrote: > >> Because I want to understand the behavior of MPI_Waitsome, I change the >> program as fellow. MPI can waitsome 10 Isend request once time, while >> waitsome 1 Irecv request each time. Why waitsome-Irecv can not be more than >> 1 as I expect? >> ===== >> MPI_Init(&argc, &argv); >> MPI_Comm_size(MPI_COMM_WORLD, &size); >> MPI_Comm_rank(MPI_COMM_WORLD, &rank); >> >> if(rank==0) { >> for(i =0; i< 10;i++){ >> MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, >> &req[i]); >> } >> printf("post send\n"); >> }else { >> for(i=0; i< 10;i++){ >> MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, >> &req[i]); >> } >> printf("post recv\n"); >> } >> >> for (;;){ >> MPI_Waitsome(10, req, &outcount, array_of_indices, stat); >> complete += outcount; >> printf("rank %d %d index %d\n", rank, outcount, >> array_of_indices[0]); >> if(complete == 10) >> break; >> } >> printf("rank %d complete\n", rank); >> MPI_Finalize(); >> ===== >> $ mpirun -n 2 ./tests >> post send >> post recv >> rank 0 10 index 0 >> rank 0 complete >> rank 1 1 index 0 >> rank 1 1 index 1 >> rank 1 1 index 2 >> rank 1 1 index 3 >> rank 1 1 index 4 >> rank 1 1 index 5 >> rank 1 1 index 6 >> rank 1 1 index 7 >> rank 1 1 index 8 >> rank 1 1 index 9 >> rank 1 complete >> >> >> >> >> >> At 2014-10-19 22:36:25, "Bland, Wesley B." <wbland(a)anl.gov> wrote: >> >> You need to move your waitany outside of the if statement so both ranks >> execute it. >> >> >> >> On Oct 19, 2014, at 9:33 AM, myself <chcdlf(a)126.com> wrote: >> >> Here is my test program >> >> ====== >> >> #include "mpi.h" >> #include <stdio.h> >> #include <stdlib.h> >> #include <sys/time.h> >> >> typedef unsigned char byte; >> >> int main(int argc, char *argv[]){ >> MPI_Init(&argc, &argv); >> int rank,size; >> MPI_Comm_size(MPI_COMM_WORLD, &size); >> MPI_Comm_rank(MPI_COMM_WORLD, &rank); >> byte buf[10][10]; >> MPI_Request req[10]; >> MPI_Status stat[10]; >> int outcount; >> int array_of_indices[10]; >> int i; >> >> if(rank==0) { >> for(i =0; i< 10;i++){ >> MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, >> &req[i]); >> } >> printf("post send\n"); >> MPI_Waitall(10, req, stat); >> printf("send over\n"); >> }else { >> for(i=0; i< 10;i++){ >> MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, >> &req[i]); >> } >> >> printf("post recv\n"); >> for (i=0;i<10;i++){ >> MPI_Waitsome(10, req, &outcount, array_of_indices, stat); >> printf("%d index %d\n", outcount, array_of_indices[0]); >> } >> } >> MPI_Finalize(); >> } >> >> ====== >> get the result like this >> ====== >> $ mpirun -n 2 ./tests >> post send >> send over >> post recv >> 1 index 0 >> 1 index 1 >> 1 index 2 >> 1 index 3 >> 1 index 4 >> 1 index 5 >> 1 index 6 >> 1 index 7 >> 1 index 8 >> 1 index 9 >> >> ====== >> >> I think I should get some completed receive connection, not just 1. >> >> >> >> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss >> >> >> >> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss >> > > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] where mpi waitsome only returns one request each time
by myself 19 Oct '14

19 Oct '14

I run several times of the program, the results are as the same as before. Could you tell me your MPICH version and compiling parameters? The version of my MPICH is 3.1.3, I compile it as $ mpirun --version HYDRA build details: Version: 3.1.3 Release Date: Wed Oct 8 09:37:19 CDT 2014 CC: gcc -fPIC CXX: g++ F77: gfortran F90: gfortran Configure options: '--disable-option-checking' '--prefix=/usr/local/mpich3' '--with-device=ch3:nemesis' 'CFLAGS=-fPIC -O2' '--disable-fortran' '--cache-file=/dev/null' '--srcdir=.' 'CC=gcc' 'LDFLAGS= ' 'LIBS=-lpthread ' 'CPPFLAGS= -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/mpl/include -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/mpl/include -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/openpa/src -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/openpa/src -D_REENTRANT -I/home/valder/Downloads/Software/mpi/mpich-3.1.3/src/mpi/romio/include' Process Manager: pmi Launchers available: ssh rsh fork slurm ll lsf sge manual persist Topology libraries available: hwloc Resource management kernels available: user slurm ll lsf sge pbs cobalt Checkpointing libraries available: Demux engines available: poll select the whole program is ========= #include "mpi.h" #include <stdio.h> #include <stdlib.h> #include <sys/time.h> typedef unsigned char byte; int main(int argc, char *argv[]){ int rank,size; byte buf[10][10]; MPI_Request req[10]; MPI_Status stat[10]; int outcount; int array_of_indices[10]; int i; int complete = 0; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); } for (;;){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); complete += outcount; printf("rank %d %d index %d\n", rank, outcount, array_of_indices[0]); if(complete == 10) break; } printf("rank %d complete\n", rank); MPI_Finalize(); } ========= I compile it as $ mpicc -o tests testsome.c and run as $ mpirun -n 2 ./tests At 2014-10-20 09:02:39, "Junchao Zhang" <jczhang(a)mcs.anl.gov> wrote: I think it is random. For example, I ran your code and its output was: post recv post send rank 0 10 index 0 rank 0 complete rank 1 1 index 0 rank 1 9 index 1 rank 1 complete --Junchao Zhang On Sun, Oct 19, 2014 at 9:48 AM, myself <chcdlf(a)126.com> wrote: Because I want to understand the behavior of MPI_Waitsome, I change the program as fellow. MPI can waitsome 10 Isend request once time, while waitsome 1 Irecv request each time. Why waitsome-Irecv can not be more than 1 as I expect? ===== MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); } for (;;){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); complete += outcount; printf("rank %d %d index %d\n", rank, outcount, array_of_indices[0]); if(complete == 10) break; } printf("rank %d complete\n", rank); MPI_Finalize(); ===== $ mpirun -n 2 ./tests post send post recv rank 0 10 index 0 rank 0 complete rank 1 1 index 0 rank 1 1 index 1 rank 1 1 index 2 rank 1 1 index 3 rank 1 1 index 4 rank 1 1 index 5 rank 1 1 index 6 rank 1 1 index 7 rank 1 1 index 8 rank 1 1 index 9 rank 1 complete At 2014-10-19 22:36:25, "Bland, Wesley B." <wbland(a)anl.gov> wrote: You need to move your waitany outside of the if statement so both ranks execute it. On Oct 19, 2014, at 9:33 AM, myself <chcdlf(a)126.com> wrote: Here is my test program ====== #include "mpi.h" #include <stdio.h> #include <stdlib.h> #include <sys/time.h> typedef unsigned char byte; int main(int argc, char *argv[]){ MPI_Init(&argc, &argv); int rank,size; MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); byte buf[10][10]; MPI_Request req[10]; MPI_Status stat[10]; int outcount; int array_of_indices[10]; int i; if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); MPI_Waitall(10, req, stat); printf("send over\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); for (i=0;i<10;i++){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); printf("%d index %d\n", outcount, array_of_indices[0]); } } MPI_Finalize(); } ====== get the result like this ====== $ mpirun -n 2 ./tests post send send over post recv 1 index 0 1 index 1 1 index 2 1 index 3 1 index 4 1 index 5 1 index 6 1 index 7 1 index 8 1 index 9 ====== I think I should get some completed receive connection, not just 1. _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] where mpi waitsome only returns one request each time
by Junchao Zhang 19 Oct '14

19 Oct '14

I think it is random. For example, I ran your code and its output was: post recv post send rank 0 10 index 0 rank 0 complete rank 1 1 index 0 rank 1 9 index 1 rank 1 complete --Junchao Zhang On Sun, Oct 19, 2014 at 9:48 AM, myself <chcdlf(a)126.com> wrote: > Because I want to understand the behavior of MPI_Waitsome, I change the > program as fellow. MPI can waitsome 10 Isend request once time, while > waitsome 1 Irecv request each time. Why waitsome-Irecv can not be more than > 1 as I expect? > ===== > MPI_Init(&argc, &argv); > MPI_Comm_size(MPI_COMM_WORLD, &size); > MPI_Comm_rank(MPI_COMM_WORLD, &rank); > > if(rank==0) { > for(i =0; i< 10;i++){ > MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, > &req[i]); > } > printf("post send\n"); > }else { > for(i=0; i< 10;i++){ > MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, > &req[i]); > } > printf("post recv\n"); > } > > for (;;){ > MPI_Waitsome(10, req, &outcount, array_of_indices, stat); > complete += outcount; > printf("rank %d %d index %d\n", rank, outcount, > array_of_indices[0]); > if(complete == 10) > break; > } > printf("rank %d complete\n", rank); > MPI_Finalize(); > ===== > $ mpirun -n 2 ./tests > post send > post recv > rank 0 10 index 0 > rank 0 complete > rank 1 1 index 0 > rank 1 1 index 1 > rank 1 1 index 2 > rank 1 1 index 3 > rank 1 1 index 4 > rank 1 1 index 5 > rank 1 1 index 6 > rank 1 1 index 7 > rank 1 1 index 8 > rank 1 1 index 9 > rank 1 complete > > > > > > At 2014-10-19 22:36:25, "Bland, Wesley B." <wbland(a)anl.gov> wrote: > > You need to move your waitany outside of the if statement so both ranks > execute it. > > > > On Oct 19, 2014, at 9:33 AM, myself <chcdlf(a)126.com> wrote: > > Here is my test program > > ====== > > #include "mpi.h" > #include <stdio.h> > #include <stdlib.h> > #include <sys/time.h> > > typedef unsigned char byte; > > int main(int argc, char *argv[]){ > MPI_Init(&argc, &argv); > int rank,size; > MPI_Comm_size(MPI_COMM_WORLD, &size); > MPI_Comm_rank(MPI_COMM_WORLD, &rank); > byte buf[10][10]; > MPI_Request req[10]; > MPI_Status stat[10]; > int outcount; > int array_of_indices[10]; > int i; > > if(rank==0) { > for(i =0; i< 10;i++){ > MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, > &req[i]); > } > printf("post send\n"); > MPI_Waitall(10, req, stat); > printf("send over\n"); > }else { > for(i=0; i< 10;i++){ > MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, > &req[i]); > } > > printf("post recv\n"); > for (i=0;i<10;i++){ > MPI_Waitsome(10, req, &outcount, array_of_indices, stat); > printf("%d index %d\n", outcount, array_of_indices[0]); > } > } > MPI_Finalize(); > } > > ====== > get the result like this > ====== > $ mpirun -n 2 ./tests > post send > send over > post recv > 1 index 0 > 1 index 1 > 1 index 2 > 1 index 3 > 1 index 4 > 1 index 5 > 1 index 6 > 1 index 7 > 1 index 8 > 1 index 9 > > ====== > > I think I should get some completed receive connection, not just 1. > > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] where mpi waitsome only returns one request each time
by myself 19 Oct '14

19 Oct '14

Because I want to understand the behavior of MPI_Waitsome, I change the program as fellow. MPI can waitsome 10 Isend request once time, while waitsome 1 Irecv request each time. Why waitsome-Irecv can not be more than 1 as I expect? ===== MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); } for (;;){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); complete += outcount; printf("rank %d %d index %d\n", rank, outcount, array_of_indices[0]); if(complete == 10) break; } printf("rank %d complete\n", rank); MPI_Finalize(); ===== $ mpirun -n 2 ./tests post send post recv rank 0 10 index 0 rank 0 complete rank 1 1 index 0 rank 1 1 index 1 rank 1 1 index 2 rank 1 1 index 3 rank 1 1 index 4 rank 1 1 index 5 rank 1 1 index 6 rank 1 1 index 7 rank 1 1 index 8 rank 1 1 index 9 rank 1 complete At 2014-10-19 22:36:25, "Bland, Wesley B." <wbland(a)anl.gov> wrote: You need to move your waitany outside of the if statement so both ranks execute it. On Oct 19, 2014, at 9:33 AM, myself <chcdlf(a)126.com> wrote: Here is my test program ====== #include "mpi.h" #include <stdio.h> #include <stdlib.h> #include <sys/time.h> typedef unsigned char byte; int main(int argc, char *argv[]){ MPI_Init(&argc, &argv); int rank,size; MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); byte buf[10][10]; MPI_Request req[10]; MPI_Status stat[10]; int outcount; int array_of_indices[10]; int i; if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); MPI_Waitall(10, req, stat); printf("send over\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); for (i=0;i<10;i++){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); printf("%d index %d\n", outcount, array_of_indices[0]); } } MPI_Finalize(); } ====== get the result like this ====== $ mpirun -n 2 ./tests post send send over post recv 1 index 0 1 index 1 1 index 2 1 index 3 1 index 4 1 index 5 1 index 6 1 index 7 1 index 8 1 index 9 ====== I think I should get some completed receive connection, not just 1. _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] where mpi waitsome only returns one request each time
by Bland, Wesley B. 19 Oct '14

19 Oct '14

You need to move your waitany outside of the if statement so both ranks execute it. On Oct 19, 2014, at 9:33 AM, myself <chcdlf(a)126.com<mailto:[email protected]>> wrote: Here is my test program ====== #include "mpi.h" #include <stdio.h> #include <stdlib.h> #include <sys/time.h> typedef unsigned char byte; int main(int argc, char *argv[]){ MPI_Init(&argc, &argv); int rank,size; MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); byte buf[10][10]; MPI_Request req[10]; MPI_Status stat[10]; int outcount; int array_of_indices[10]; int i; if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); MPI_Waitall(10, req, stat); printf("send over\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); for (i=0;i<10;i++){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); printf("%d index %d\n", outcount, array_of_indices[0]); } } MPI_Finalize(); } ====== get the result like this ====== $ mpirun -n 2 ./tests post send send over post recv 1 index 0 1 index 1 1 index 2 1 index 3 1 index 4 1 index 5 1 index 6 1 index 7 1 index 8 1 index 9 ====== I think I should get some completed receive connection, not just 1. _______________________________________________ discuss mailing list discuss(a)mpich.org<mailto:[email protected]> To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

where mpi waitsome only returns one request each time
by myself 19 Oct '14

19 Oct '14

Here is my test program ====== #include "mpi.h" #include <stdio.h> #include <stdlib.h> #include <sys/time.h> typedef unsigned char byte; int main(int argc, char *argv[]){ MPI_Init(&argc, &argv); int rank,size; MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); byte buf[10][10]; MPI_Request req[10]; MPI_Status stat[10]; int outcount; int array_of_indices[10]; int i; if(rank==0) { for(i =0; i< 10;i++){ MPI_Isend(buf[i], 10, MPI_BYTE, 1, 123, MPI_COMM_WORLD, &req[i]); } printf("post send\n"); MPI_Waitall(10, req, stat); printf("send over\n"); }else { for(i=0; i< 10;i++){ MPI_Irecv(buf[i], 10, MPI_BYTE, 0, 123, MPI_COMM_WORLD, &req[i]); } printf("post recv\n"); for (i=0;i<10;i++){ MPI_Waitsome(10, req, &outcount, array_of_indices, stat); printf("%d index %d\n", outcount, array_of_indices[0]); } } MPI_Finalize(); } ====== get the result like this ====== $ mpirun -n 2 ./tests post send send over post recv 1 index 0 1 index 1 1 index 2 1 index 3 1 index 4 1 index 5 1 index 6 1 index 7 1 index 8 1 index 9 ====== I think I should get some completed receive connection, not just 1. _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] aggregator in collective I/O
by Jaln 17 Oct '14

17 Oct '14

Thanks Rob very much, Jialin On Fri, Oct 17, 2014 at 9:45 AM, Rob Latham <robl(a)mcs.anl.gov> wrote: > > > On 10/16/2014 10:07 PM, Jaln wrote: > >> Hi, >> How can I know the aggregator's rank in the collective I/O? >> > > fd->hints->ranklist[] contains an array of MPI ranks assigned to an > aggregator. > > > I also feel difficult to distinguish the aggregator and non-aggregator >> in the ad_read_coll.c, >> for example, in function ADIOI_R_Exchange_data, >> how can I know it is executed by a process or an aggregator? >> I can only see how 'nprocs_for_coll' is used in partitioning the file >> domain, >> but couldn't see how it is used to tell the certain number of aggregator >> to do the collective I/O. >> > > I think you want to look at ADIOI_Calc_aggregator, which will report which > MPI-IO aggregator is responsible for which regions of a file, and how much. > > ADIOI_R_Exchange_data is called by every process. Some processes are > receiving data from an aggregator, and aggregators are sending read data, > but everyone participates. > > Take a look earlier on: ADIOI_Calc_my_req and ADIOI_Calc_others_req are > two of the routines responsible for generating the schedule of requests. > > ==rob > > -- > Rob Latham > Mathematics and Computer Science Division > Argonne National Lab, IL USA > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] aggregator in collective I/O
by Rob Latham 17 Oct '14

17 Oct '14

On 10/16/2014 10:07 PM, Jaln wrote: > Hi, > How can I know the aggregator's rank in the collective I/O? fd->hints->ranklist[] contains an array of MPI ranks assigned to an aggregator. > I also feel difficult to distinguish the aggregator and non-aggregator > in the ad_read_coll.c, > for example, in function ADIOI_R_Exchange_data, > how can I know it is executed by a process or an aggregator? > I can only see how 'nprocs_for_coll' is used in partitioning the file > domain, > but couldn't see how it is used to tell the certain number of aggregator > to do the collective I/O. I think you want to look at ADIOI_Calc_aggregator, which will report which MPI-IO aggregator is responsible for which regions of a file, and how much. ADIOI_R_Exchange_data is called by every process. Some processes are receiving data from an aggregator, and aggregators are sending read data, but everyone participates. Take a look earlier on: ADIOI_Calc_my_req and ADIOI_Calc_others_req are two of the routines responsible for generating the schedule of requests. ==rob -- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

aggregator in collective I/O
by Jaln 16 Oct '14

16 Oct '14

Hi, How can I know the aggregator's rank in the collective I/O? I also feel difficult to distinguish the aggregator and non-aggregator in the ad_read_coll.c, for example, in function ADIOI_R_Exchange_data, how can I know it is executed by a process or an aggregator? I can only see how 'nprocs_for_coll' is used in partitioning the file domain, but couldn't see how it is used to tell the certain number of aggregator to do the collective I/O. Thanks, Jialin _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] Problem with RMA && MPI_Type_vector when stride value is relatively large (UPD)
by Zhao, Xin 16 Oct '14

16 Oct '14

Hi Victor, I think it is due to a bug for MPI_PUT to handle large derived datatype and we will fix it as soon as possible. The code works with MPI_GET. You can track this bug at: https://trac.mpich.org/projects/mpich/ticket/2189 Thanks, Xin ________________________________ From: Victor Vysotskiy [victor.vysotskiy(a)teokem.lu.se] Sent: Thursday, October 16, 2014 11:19 AM To: discuss(a)mpich.org Subject: Re: [mpich-discuss] Problem with RMA && MPI_Type_vector when stride value is relatively large (UPD) Hi, I am sorry that the test bed code has not been attached to the original message. Enclosed please find it. I have also checked the issue in MPICH v3.1.3. Unfortunately, the latest MPICH3 version behaves exactly the same as the MPICH2 v1.5, i.e. it also hangs at "30000 2000". With best regards, Victor. P.s. Below is given an update on MPICH3 version used: MPICH2 Version: 3.1.3 MPICH2 Release date: Wed Oct 8 09:37:19 CDT 2014 MPICH2 Device: ch3:nemesis MPICH2 configure: --prefix=/opt/mpi/mpich3-1.3/ --disable-weak-symbols MPICH2 CC: gcc -O2 MPICH2 CXX: c++ -O2 MPICH2 F77: gfortran -O2 MPICH2 FC: gfortran -O2 GCC version 4.7.3 ________________________________ From: Victor Vysotskiy Sent: Thursday, October 16, 2014 1:34 PM To: discuss(a)mpich.org Subject: Problem with RMA && MPI_Type_vector when stride value is relatively large Hi, I am straggling with a simple MPI program when a vector derived datatype is used as origin/remote datatypes in the MPI_Put RMA operation. The underlying idea is quite simple: a contiguous piece of memory is interpreted as a 2D array and only its certain slice is going to be updated via MPI_Put. To this end, I am using a MPI_Type_vector call. Everything works fine when the stride value <= 24000, regardless count and blocklength. However, if stride is greater than 24000 then program simply hangs. The source code is attached and you can simply verify the issue by compiling and running it with the following parameters: mpicc mpi_tvec_rma.c mpirun -np 2 ./a.out 10000 2000 Ok mpirun -np 2 ./a.out 20000 2000 Ok mpirun -np 2 ./a.out 30000 2000 Hangs Is there any internal/known limits to the size of MPI_TYPE_VECTOR datatypes? So, I am really puzzled what is happening here and any help will be greatly appreciated! With best regards, Victor. P.S. The attached code is perfectly runs with the OpenMPI v1.8.2. P.P.S. Below are given some details about hardware and software setups used: MPICH2 Version: 1.5 MPICH2 Release date: Mon Oct 8 14:00:48 CDT 2012 MPICH2 Device: ch3:nemesis MPICH2 configure: --prefix=/opt/mpi/mpich2/ --disable-weak-symbols MPICH2 CC: gcc -O2 MPICH2 CXX: c++ -O2 MPICH2 F77: gfortran -O2 MPICH2 FC: gfortran -O2 GCC version 4.7.3 _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0