A customer (Argonne ;) is complaining about O(p) allocations in collective i/o.  A collective read is failing at larger scale.  

Any thoughts or comments or advice?   There appears to be lots of O(p) in ROMIO collective I/O.  Plus a lot of (possibly large) aggregated data buffers.  A quick search shows

The common ROMIO read collective code:

Find all "ADIOI_Malloc", Match case, Regular expression (UNIX)
File Z:\bgq\comm\lib\dev\mpich2\src\mpi\romio\adio\common\ad_read_coll.c
  124 38:        st_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset));
  125 39:        end_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset));
  317 44:        *offset_list_ptr = (ADIO_Offset *) ADIOI_Malloc(2*sizeof(ADIO_Offset));
  318 41:        *len_list_ptr = (ADIO_Offset *) ADIOI_Malloc(2*sizeof(ADIO_Offset));
  334 44:        *offset_list_ptr = (ADIO_Offset *) ADIOI_Malloc(2*sizeof(ADIO_Offset));
  335 41:        *len_list_ptr = (ADIO_Offset *) ADIOI_Malloc(2*sizeof(ADIO_Offset));
  436 18:                 ADIOI_Malloc((contig_access_count+1)*sizeof(ADIO_Offset));  
  437 41:        *len_list_ptr = (ADIO_Offset *) ADIOI_Malloc((contig_access_count+1)*sizeof(ADIO_Offset));
  573 37:    if (ntimes) read_buf = (char *) ADIOI_Malloc(coll_bufsize);
  578 21:    count = (int *) ADIOI_Malloc(nprocs * sizeof(int));
  587 25:    send_size = (int *) ADIOI_Malloc(nprocs * sizeof(int));
  590 25:    recv_size = (int *) ADIOI_Malloc(nprocs * sizeof(int));
  598 25:    start_pos = (int *) ADIOI_Malloc(nprocs*sizeof(int));
  739 32:            tmp_buf = (char *) ADIOI_Malloc(for_next_iter);
  744 33:            read_buf = (char *) ADIOI_Malloc(for_next_iter+coll_bufsize);
  805 9:        ADIOI_Malloc((nprocs_send+nprocs_recv+1)*sizeof(MPI_Request));
  827 30:        recv_buf = (char **) ADIOI_Malloc(nprocs * sizeof(char*));
  830 44:                                  (char *) ADIOI_Malloc(recv_size[i]);
  870 31:    statuses = (MPI_Status *) ADIOI_Malloc((nprocs_send+nprocs_recv+1) * \
  988 35:    curr_from_proc = (unsigned *) ADIOI_Malloc(nprocs * sizeof(unsigned));
  989 35:    done_from_proc = (unsigned *) ADIOI_Malloc(nprocs * sizeof(unsigned));
  990 35:    recv_buf_idx   = (unsigned *) ADIOI_Malloc(nprocs * sizeof(unsigned));
Total found: 22

Our BG version of read collective:

File Z:\bgq\comm\lib\dev\mpich2\src\mpi\romio\adio\ad_bg\ad_bg_rdcoll.c
  179 40:        st_offsets   = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset));
  180 40:        end_offsets  = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset));
  183 43:            bg_offsets0 = (ADIO_Offset *) ADIOI_Malloc(2*nprocs*sizeof(ADIO_Offset));
  184 43:            bg_offsets  = (ADIO_Offset *) ADIOI_Malloc(2*nprocs*sizeof(ADIO_Offset));
  475 37:    if (ntimes) read_buf = (char *) ADIOI_Malloc(coll_bufsize);
  480 21:    count = (int *) ADIOI_Malloc(nprocs * sizeof(int));
  489 25:    send_size = (int *) ADIOI_Malloc(nprocs * sizeof(int));
  492 25:    recv_size = (int *) ADIOI_Malloc(nprocs * sizeof(int));
  500 25:    start_pos = (int *) ADIOI_Malloc(nprocs*sizeof(int));
  676 32:            tmp_buf = (char *) ADIOI_Malloc(for_next_iter);
  681 33:            read_buf = (char *) ADIOI_Malloc(for_next_iter+coll_bufsize);
  761 9:        ADIOI_Malloc((nprocs_send+nprocs_recv+1)*sizeof(MPI_Request));
  783 30:        recv_buf = (char **) ADIOI_Malloc(nprocs * sizeof(char*));
  786 44:                                  (char *) ADIOI_Malloc(recv_size[i]);
  826 31:    statuses = (MPI_Status *) ADIOI_Malloc((nprocs_send+nprocs_recv+1) * \
  944 35:    curr_from_proc = (unsigned *) ADIOI_Malloc(nprocs * sizeof(unsigned));
  945 35:    done_from_proc = (unsigned *) ADIOI_Malloc(nprocs * sizeof(unsigned));
  946 35:    recv_buf_idx   = (unsigned *) ADIOI_Malloc(nprocs * sizeof(unsigned));
  1058 23:    rdispls = (int *) ADIOI_Malloc( nprocs * sizeof(int) );
  1063 29:    all_recv_buf = (char *) ADIOI_Malloc( rtail );
  1064 26:    recv_buf = (char **) ADIOI_Malloc(nprocs * sizeof(char *));
  1068 23:    sdispls = (int *) ADIOI_Malloc( nprocs * sizeof(int) );
  1073 29:    all_send_buf = (char *) ADIOI_Malloc( stail );
Total found: 23


Bob Cernohous:  (T/L 553) 507-253-6093

BobC@us.ibm.com
IBM Rochester, Building 030-2(C335), Department 61L
3605 Hwy 52 North, Rochester,  MN 55901-7829

> Chaos reigns within.
> Reflect, repent, and reboot.
> Order shall return.