📄 ad_bgl_wrcoll.c
字号:
/* ---------------------------------------------------------------- *//* (C)Copyright IBM Corp. 2007, 2008 *//* ---------------------------------------------------------------- *//** * \file ad_bgl_wrcoll.c * \brief ??? *//* -*- Mode: C; c-basic-offset:4 ; -*- *//* * Copyright (C) 1997 University of Chicago. * See COPYRIGHT notice in top-level directory. */#include "adio.h"#include "adio_extern.h"#include "ad_bgl.h"#include "ad_bgl_pset.h"#include "ad_bgl_aggrs.h"#ifdef PROFILE#include "mpe.h"#endif/* prototypes of functions used for collective writes only. */static void ADIOI_Exch_and_write(ADIO_File fd, void *buf, MPI_Datatype datatype, int nprocs, int myrank, ADIOI_Access *others_req, ADIO_Offset *offset_list, int *len_list, int contig_access_count, ADIO_Offset min_st_offset, ADIO_Offset fd_size, ADIO_Offset *fd_start, ADIO_Offset *fd_end, int *buf_idx, int *error_code);static void ADIOI_W_Exchange_data(ADIO_File fd, void *buf, char *write_buf, ADIOI_Flatlist_node *flat_buf, ADIO_Offset *offset_list, int *len_list, int *send_size, int *recv_size, ADIO_Offset off, int size, int *count, int *start_pos, int *partial_recv, int *sent_to_proc, int nprocs, int myrank, int buftype_is_contig, int contig_access_count, ADIO_Offset min_st_offset, ADIO_Offset fd_size, ADIO_Offset *fd_start, ADIO_Offset *fd_end, ADIOI_Access *others_req, int *send_buf_idx, int *curr_to_proc, int *done_to_proc, int *hole, int iter, MPI_Aint buftype_extent, int *buf_idx, int *error_code);static void ADIOI_W_Exchange_data_alltoallv( ADIO_File fd, void *buf, char *write_buf, /* 1 */ ADIOI_Flatlist_node *flat_buf, ADIO_Offset *offset_list, int *len_list, int *send_size, int *recv_size, ADIO_Offset off, int size, /* 2 */ int *count, int *start_pos, int *partial_recv, int *sent_to_proc, int nprocs, int myrank, int buftype_is_contig, int contig_access_count, ADIO_Offset min_st_offset, ADIO_Offset fd_size, ADIO_Offset *fd_start, ADIO_Offset *fd_end, ADIOI_Access *others_req, int *send_buf_idx, int *curr_to_proc, /* 3 */ int *done_to_proc, int *hole, /* 4 */ int iter, MPI_Aint buftype_extent, int *buf_idx, int *error_code);static void ADIOI_Fill_send_buffer(ADIO_File fd, void *buf, ADIOI_Flatlist_node *flat_buf, char **send_buf, ADIO_Offset *offset_list, int *len_list, int *send_size, MPI_Request *requests, int *sent_to_proc, int nprocs, int myrank, int contig_access_count, ADIO_Offset min_st_offset, ADIO_Offset fd_size, ADIO_Offset *fd_start, ADIO_Offset *fd_end, int *send_buf_idx, int *curr_to_proc, int *done_to_proc, int iter, MPI_Aint buftype_extent);static void ADIOI_Fill_send_buffer_nosend(ADIO_File fd, void *buf, ADIOI_Flatlist_node *flat_buf, char **send_buf, ADIO_Offset *offset_list, int *len_list, int *send_size, MPI_Request *requests, int *sent_to_proc, int nprocs, int myrank, int contig_access_count, ADIO_Offset min_st_offset, ADIO_Offset fd_size, ADIO_Offset *fd_start, ADIO_Offset *fd_end, int *send_buf_idx, int *curr_to_proc, int *done_to_proc, int iter, MPI_Aint buftype_extent);static void ADIOI_Heap_merge(ADIOI_Access *others_req, int *count, ADIO_Offset *srt_off, int *srt_len, int *start_pos, int nprocs, int nprocs_recv, int total_elements);void ADIOI_BGL_WriteStridedColl(ADIO_File fd, void *buf, int count, MPI_Datatype datatype, int file_ptr_type, ADIO_Offset offset, ADIO_Status *status, int *error_code){/* Uses a generalized version of the extended two-phase method described in "An Extended Two-Phase Method for Accessing Sections of Out-of-Core Arrays", Rajeev Thakur and Alok Choudhary, Scientific Programming, (5)4:301--317, Winter 1996. http://www.mcs.anl.gov/home/thakur/ext2ph.ps */ ADIOI_Access *my_req; /* array of nprocs access structures, one for each other process in whose file domain this process's request lies */ ADIOI_Access *others_req; /* array of nprocs access structures, one for each other process whose request lies in this process's file domain. */ int i, filetype_is_contig, nprocs, nprocs_for_coll, myrank; int contig_access_count=0, interleave_count = 0, buftype_is_contig; int *count_my_req_per_proc, count_my_req_procs, count_others_req_procs; ADIO_Offset orig_fp, start_offset, end_offset, fd_size, min_st_offset, off; ADIO_Offset *offset_list = NULL, *st_offsets = NULL, *fd_start = NULL, *fd_end = NULL, *end_offsets = NULL; ADIO_Offset *bgl_offsets0 = NULL, *bgl_offsets = NULL; int ii; int *buf_idx = NULL, *len_list = NULL; double io_time = 0, all_time, max_all_time; double tstep1, max_tstep1; double tstep1_1, max_tstep1_1; double tstep1_2, max_tstep1_2; double tstep1_3, max_tstep1_3; double tstep2, max_tstep2; double tstep3, max_tstep3; double tstep4, max_tstep4; double sum_sz;#if BGL_PROFILE BGLMPIO_T_CIO_RESET( 0, w )#endif#ifdef PROFILE MPE_Log_event(13, 0, "start computation");#endif MPI_Comm_size(fd->comm, &nprocs); MPI_Comm_rank(fd->comm, &myrank);/* the number of processes that actually perform I/O, nprocs_for_coll, * is stored in the hints off the ADIO_File structure */ nprocs_for_coll = fd->hints->cb_nodes; orig_fp = fd->fp_ind;#if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 0, 1, 0, BGLMPIO_CIO_LCOMP, BGLMPIO_CIO_LAST )#endif /* only check for interleaving if cb_write isn't disabled */ if (fd->hints->cb_write != ADIOI_HINT_DISABLE) { /* For this process's request, calculate the list of offsets and lengths in the file and determine the start and end offsets. */ /* Note: end_offset points to the last byte-offset that will be accessed. e.g., if start_offset=0 and 100 bytes to be read, end_offset=99*/ ADIOI_Calc_my_off_len(fd, count, datatype, file_ptr_type, offset, &offset_list, &len_list, &start_offset, &end_offset, &contig_access_count); #if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 1, 1, 1, BGLMPIO_CIO_GATHER, BGLMPIO_CIO_LCOMP )#endif /* each process communicates its start and end offsets to other processes. The result is an array each of start and end offsets stored in order of process rank. */ st_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset)); end_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset)); if (bglmpio_tunegather) { bgl_offsets0 = (ADIO_Offset *) ADIOI_Malloc(2*nprocs*sizeof(ADIO_Offset)); bgl_offsets = (ADIO_Offset *) ADIOI_Malloc(2*nprocs*sizeof(ADIO_Offset)); for (ii=0; ii<nprocs; ii++) { bgl_offsets0[ii*2] = 0; bgl_offsets0[ii*2+1] = 0; } bgl_offsets0[myrank*2] = start_offset; bgl_offsets0[myrank*2+1] = end_offset; MPI_Allreduce( bgl_offsets0, bgl_offsets, nprocs*2, ADIO_OFFSET, MPI_MAX, fd->comm ); for (ii=0; ii<nprocs; ii++) { st_offsets [ii] = bgl_offsets[ii*2] ; end_offsets[ii] = bgl_offsets[ii*2+1]; } ADIOI_Free( bgl_offsets0 ); ADIOI_Free( bgl_offsets ); } else { MPI_Allgather(&start_offset, 1, ADIO_OFFSET, st_offsets, 1, ADIO_OFFSET, fd->comm); MPI_Allgather(&end_offset, 1, ADIO_OFFSET, end_offsets, 1, ADIO_OFFSET, fd->comm); }#if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 0, 1, 1, BGLMPIO_CIO_PATANA, BGLMPIO_CIO_GATHER )#endif /* are the accesses of different processes interleaved? */ for (i=1; i<nprocs; i++) if ((st_offsets[i] < end_offsets[i-1]) && (st_offsets[i] <= end_offsets[i])) interleave_count++; /* This is a rudimentary check for interleaving, but should suffice for the moment. */ } ADIOI_Datatype_iscontig(datatype, &buftype_is_contig); if (fd->hints->cb_write == ADIOI_HINT_DISABLE || (!interleave_count && (fd->hints->cb_write == ADIOI_HINT_AUTO))) { /* use independent accesses */ if (fd->hints->cb_write != ADIOI_HINT_DISABLE) { ADIOI_Free(offset_list); ADIOI_Free(len_list); ADIOI_Free(st_offsets); ADIOI_Free(end_offsets); } fd->fp_ind = orig_fp; ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig); if (buftype_is_contig && filetype_is_contig) { if (file_ptr_type == ADIO_EXPLICIT_OFFSET) { off = fd->disp + (fd->etype_size) * offset; ADIO_WriteContig(fd, buf, count, datatype, ADIO_EXPLICIT_OFFSET, off, status, error_code); } else ADIO_WriteContig(fd, buf, count, datatype, ADIO_INDIVIDUAL, 0, status, error_code); } else ADIO_WriteStrided(fd, buf, count, datatype, file_ptr_type, offset, status, error_code); return; }#if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 1, 1, 1, BGLMPIO_CIO_FD_PART, BGLMPIO_CIO_PATANA )#endif /* Divide the I/O workload among "nprocs_for_coll" processes. This is done by (logically) dividing the file into file domains (FDs); each process may directly access only its own file domain. */ if (bglmpio_tuneblocking) ADIOI_BGL_GPFS_Calc_file_domains(st_offsets, end_offsets, nprocs, nprocs_for_coll, &min_st_offset, &fd_start, &fd_end, &fd_size, fd->fs_ptr); else ADIOI_Calc_file_domains(st_offsets, end_offsets, nprocs, nprocs_for_coll, &min_st_offset, &fd_start, &fd_end, &fd_size); #if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 0, 1, 1, BGLMPIO_CIO_MYREQ, BGLMPIO_CIO_FD_PART )#endif /* calculate what portions of the access requests of this process are located in what file domains */ if (bglmpio_tuneblocking) ADIOI_BGL_Calc_my_req(fd, offset_list, len_list, contig_access_count, min_st_offset, fd_start, fd_end, fd_size, nprocs, &count_my_req_procs, &count_my_req_per_proc, &my_req, &buf_idx); else ADIOI_Calc_my_req(fd, offset_list, len_list, contig_access_count, min_st_offset, fd_start, fd_end, fd_size, nprocs, &count_my_req_procs, &count_my_req_per_proc, &my_req, &buf_idx); #if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 1, 1, 1, BGLMPIO_CIO_OTHREQ, BGLMPIO_CIO_MYREQ )#endif /* based on everyone's my_req, calculate what requests of other processes lie in this process's file domain. count_others_req_procs = number of processes whose requests lie in this process's file domain (including this process itself) count_others_req_per_proc[i] indicates how many separate contiguous requests of proc. i lie in this process's file domain. */ if (bglmpio_tuneblocking) ADIOI_BGL_Calc_others_req(fd, count_my_req_procs, count_my_req_per_proc, my_req, nprocs, myrank, &count_others_req_procs, &others_req); else ADIOI_Calc_others_req(fd, count_my_req_procs, count_my_req_per_proc, my_req, nprocs, myrank, &count_others_req_procs, &others_req); #if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 1, 1, 1, BGLMPIO_CIO_DEXCH, BGLMPIO_CIO_OTHREQ )#endif ADIOI_Free(count_my_req_per_proc); for (i=0; i < nprocs; i++) { if (my_req[i].count) { ADIOI_Free(my_req[i].offsets); ADIOI_Free(my_req[i].lens); } } ADIOI_Free(my_req);/* exchange data and write in sizes of no more than coll_bufsize. */ ADIOI_Exch_and_write(fd, buf, datatype, nprocs, myrank, others_req, offset_list, len_list, contig_access_count, min_st_offset, fd_size, fd_start, fd_end, buf_idx, error_code);#if BGL_PROFILE BGLMPIO_T_CIO_SET_GET( 0, w, 1, 0, 1, BGLMPIO_CIO_LAST, BGLMPIO_CIO_T_DEXCH ) BGLMPIO_T_CIO_SET_GET( 0, w, 0, 0, 1, BGLMPIO_CIO_LAST, BGLMPIO_CIO_T_MPIO_CRW ) BGLMPIO_T_CIO_REPORT( 0, w, fd, myrank )#endif /* free all memory allocated for collective I/O */ for (i=0; i<nprocs; i++) { if (others_req[i].count) { ADIOI_Free(others_req[i].offsets); ADIOI_Free(others_req[i].lens); ADIOI_Free(others_req[i].mem_ptrs); } } ADIOI_Free(others_req); ADIOI_Free(buf_idx); ADIOI_Free(offset_list); ADIOI_Free(len_list); ADIOI_Free(st_offsets); ADIOI_Free(end_offsets); ADIOI_Free(fd_start); ADIOI_Free(fd_end);#ifdef HAVE_STATUS_SET_BYTES if (status) { int bufsize, size; /* Don't set status if it isn't needed */ MPI_Type_size(datatype, &size); bufsize = size * count; MPIR_Status_set_bytes(status, datatype, bufsize); }/* This is a temporary way of filling in status. The right way is to keep track of how much data was actually written during collective I/O. */#endif
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -