Home Research Papers Posters Manuals Talks Download People Help Internal

libs/ck-libs/ampi/romio/adio/ad_lustre/ad_lustre_aggregate.c

Go to the documentation of this file.

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
 *   Copyright (C) 1997 University of Chicago.
 *   See COPYRIGHT notice in top-level directory.
 *
 *   Copyright (C) 2007 Oak Ridge National Laboratory
 *
 *   Copyright (C) 2008 Sun Microsystems, Lustre group
 */

#include "ad_lustre.h"
#include "adio_extern.h"

#undef AGG_DEBUG

void ADIOI_LUSTRE_Get_striping_info(ADIO_File fd, int **striping_info_ptr,
                    int mode)
{
    int *striping_info = NULL;
    /* get striping information:
     *  striping_info[0]: stripe_size
     *  striping_info[1]: stripe_count
     *  striping_info[2]: avail_cb_nodes
     */
    int stripe_size, stripe_count, CO = 1;
    int avail_cb_nodes, divisor, nprocs_for_coll = fd->hints->cb_nodes;

    /* Get hints value */
    /* stripe size */
    stripe_size = fd->hints->striping_unit;
    /* stripe count */
    /* stripe_size and stripe_count have been validated in ADIOI_LUSTRE_Open() */
    stripe_count = fd->hints->striping_factor;

    /* Calculate the available number of I/O clients */
    if (!mode) {
        /* for collective read,
     * if "CO" clients access the same OST simultaneously,
     * the OST disk seek time would be much. So, to avoid this,
     * it might be better if 1 client only accesses 1 OST.
     * So, we set CO = 1 to meet the above requirement.
     */
    CO = 1;
    /*XXX: maybe there are other better way for collective read */
    } else {
        /* CO also has been validated in ADIOI_LUSTRE_Open(), >0 */
    CO = fd->hints->fs_hints.lustre.co_ratio;
    }
    /* Calculate how many IO clients we need */
    /* Algorithm courtesy Pascal Deveze (pascal.deveze@bull.net) */
    /* To avoid extent lock conflicts,
     * avail_cb_nodes should either 
     *  - be a multiple of stripe_count,
     *  - or divide stripe_count exactly
     * so that each OST is accessed by a maximum of CO constant clients. */
    if (nprocs_for_coll >= stripe_count)
    /* avail_cb_nodes should be a multiple of stripe_count and the number
     * of procs per OST should be limited to the minimum between
     * nprocs_for_coll/stripe_count and CO 
     * 
     * e.g. if stripe_count=20, nprocs_for_coll=42 and CO=3 then 
     * avail_cb_nodes should be equal to 40 */
        avail_cb_nodes = 
        stripe_count * ADIOI_MIN(nprocs_for_coll/stripe_count, CO);
    else {
        /* nprocs_for_coll is less than stripe_count */
        /* avail_cb_nodes should divide stripe_count */
        /* e.g. if stripe_count=60 and nprocs_for_coll=8 then 
     * avail_cb_nodes should be egal to 6 */
        /* This could be done with :
                while (stripe_count % avail_cb_nodes != 0) avail_cb_nodes--;
       but this can be optimized for large values of nprocs_for_coll and
       stripe_count */
        divisor = 2;
        avail_cb_nodes = 1;
        /* try to divise */
        while (stripe_count >= divisor*divisor) {
            if ((stripe_count % divisor) == 0) {
                 if (stripe_count/divisor <= nprocs_for_coll) {
                     /* The value is found ! */
                     avail_cb_nodes = stripe_count/divisor;
                     break;
        }
        /* if divisor is less than nprocs_for_coll, divisor is a
         * solution, but it is not sure that it is the best one */
                else if (divisor <= nprocs_for_coll) 
            avail_cb_nodes = divisor;
        }
        divisor++;
        }
    }

    *striping_info_ptr = (int *) ADIOI_Malloc(3 * sizeof(int));
    striping_info = *striping_info_ptr;
    striping_info[0] = stripe_size;
    striping_info[1] = stripe_count;
    striping_info[2] = avail_cb_nodes;
}

int ADIOI_LUSTRE_Calc_aggregator(ADIO_File fd, ADIO_Offset off,
                                 ADIO_Offset *len, int *striping_info)
{
    int rank_index, rank;
    ADIO_Offset avail_bytes;
    int stripe_size = striping_info[0];
    int avail_cb_nodes = striping_info[2];

    /* Produce the stripe-contiguous pattern for Lustre */
    rank_index = (int)((off / stripe_size) % avail_cb_nodes);

    /* we index into fd_end with rank_index, and fd_end was allocated to be no
     * bigger than fd->hins->cb_nodes.   If we ever violate that, we're
     * overrunning arrays.  Obviously, we should never ever hit this abort
     */
    if (rank_index >= fd->hints->cb_nodes)
        MPI_Abort(MPI_COMM_WORLD, 1);

    avail_bytes = (off / (ADIO_Offset)stripe_size + 1) *
                  (ADIO_Offset)stripe_size - off;
    if (avail_bytes < *len) {
    /* this proc only has part of the requested contig. region */
    *len = avail_bytes;
    }
    /* map our index to a rank */
    /* NOTE: FOR NOW WE DON'T HAVE A MAPPING...JUST DO 0..NPROCS_FOR_COLL */
    rank = fd->hints->ranklist[rank_index];

    return rank;
}

/* ADIOI_LUSTRE_Calc_my_req() - calculate what portions of the access requests
 * of this process are located in the file domains of various processes
 * (including this one)
 */


void ADIOI_LUSTRE_Calc_my_req(ADIO_File fd, ADIO_Offset *offset_list,
                  ADIO_Offset *len_list, int contig_access_count,
                  int *striping_info, int nprocs,
                              int *count_my_req_procs_ptr,
                  int **count_my_req_per_proc_ptr,
                  ADIOI_Access **my_req_ptr,
                  int ***buf_idx_ptr)
{
    /* Nothing different from ADIOI_Calc_my_req(), except calling
     * ADIOI_Lustre_Calc_aggregator() instead of the old one */
    int *count_my_req_per_proc, count_my_req_procs, **buf_idx;
    int i, l, proc;
    ADIO_Offset avail_len, rem_len, curr_idx, off;
    ADIOI_Access *my_req;

    *count_my_req_per_proc_ptr = (int *) ADIOI_Calloc(nprocs, sizeof(int));
    count_my_req_per_proc = *count_my_req_per_proc_ptr;
    /* count_my_req_per_proc[i] gives the no. of contig. requests of this
     * process in process i's file domain. calloc initializes to zero.
     * I'm allocating memory of size nprocs, so that I can do an
     * MPI_Alltoall later on.
     */

    buf_idx = (int **) ADIOI_Malloc(nprocs * sizeof(int*));

    /* one pass just to calculate how much space to allocate for my_req;
     * contig_access_count was calculated way back in ADIOI_Calc_my_off_len()
     */
    for (i = 0; i < contig_access_count; i++) {
    /* short circuit offset/len processing if len == 0
     * (zero-byte  read/write
     */
    if (len_list[i] == 0)
        continue;
    off = offset_list[i];
    avail_len = len_list[i];
    /* note: we set avail_len to be the total size of the access.
     * then ADIOI_LUSTRE_Calc_aggregator() will modify the value to return
     * the amount that was available.
     */
    proc = ADIOI_LUSTRE_Calc_aggregator(fd, off, &avail_len, striping_info);
    count_my_req_per_proc[proc]++;

    /* figure out how many data is remaining in the access
     * we'll take care of this data (if there is any)
     * in the while loop below.
     */
    rem_len = len_list[i] - avail_len;

    while (rem_len != 0) {
        off += avail_len;   /* point to first remaining byte */
        avail_len = rem_len;    /* save remaining size, pass to calc */
        proc = ADIOI_LUSTRE_Calc_aggregator(fd, off, &avail_len, striping_info);
        count_my_req_per_proc[proc]++;
        rem_len -= avail_len;   /* reduce remaining length by amount from fd */
    }
    }

    /* buf_idx is relevant only if buftype_is_contig.
     * buf_idx[i] gives the index into user_buf where data received
     * from proc 'i' should be placed. This allows receives to be done
     * without extra buffer. This can't be done if buftype is not contig.
     */

    /* initialize buf_idx vectors */
    for (i = 0; i < nprocs; i++) {
    /* add one to count_my_req_per_proc[i] to avoid zero size malloc */
    buf_idx[i] = (int *) ADIOI_Malloc((count_my_req_per_proc[i] + 1)
                               * sizeof(int)); 
    }

    /* now allocate space for my_req, offset, and len */
    *my_req_ptr = (ADIOI_Access *) ADIOI_Malloc(nprocs * sizeof(ADIOI_Access));
    my_req = *my_req_ptr;

    count_my_req_procs = 0;
    for (i = 0; i < nprocs; i++) {
    if (count_my_req_per_proc[i]) {
        my_req[i].offsets = (ADIO_Offset *)
                        ADIOI_Malloc(count_my_req_per_proc[i] *
                                             sizeof(ADIO_Offset));
        my_req[i].lens = (int *) ADIOI_Malloc(count_my_req_per_proc[i] *
                                  sizeof(int));
        count_my_req_procs++;
    }
    my_req[i].count = 0;    /* will be incremented where needed later */
    }

    /* now fill in my_req */
    curr_idx = 0;
    for (i = 0; i < contig_access_count; i++) {
    /* short circuit offset/len processing if len == 0
     *  (zero-byte  read/write */
    if (len_list[i] == 0)
        continue;
    off = offset_list[i];
    avail_len = len_list[i];
    proc = ADIOI_LUSTRE_Calc_aggregator(fd, off, &avail_len, striping_info);

    l = my_req[proc].count;

    ADIOI_Assert(curr_idx == (int) curr_idx);
    ADIOI_Assert(l < count_my_req_per_proc[proc]);
    buf_idx[proc][l] = (int) curr_idx;
    curr_idx += avail_len;

    rem_len = len_list[i] - avail_len;

    /* store the proc, offset, and len information in an array
     * of structures, my_req. Each structure contains the
     * offsets and lengths located in that process's FD,
     * and the associated count.
     */
    my_req[proc].offsets[l] = off;
    ADIOI_Assert(avail_len == (int) avail_len);
    my_req[proc].lens[l] = (int) avail_len;
    my_req[proc].count++;

    while (rem_len != 0) {
        off += avail_len;
        avail_len = rem_len;
        proc = ADIOI_LUSTRE_Calc_aggregator(fd, off, &avail_len,
                                                striping_info);

        l = my_req[proc].count;
        ADIOI_Assert(curr_idx == (int) curr_idx);
        ADIOI_Assert(l < count_my_req_per_proc[proc]);
        buf_idx[proc][l] = (int) curr_idx;

        curr_idx += avail_len;
        rem_len -= avail_len;

        my_req[proc].offsets[l] = off;
        ADIOI_Assert(avail_len == (int) avail_len);
        my_req[proc].lens[l] = (int) avail_len;
        my_req[proc].count++;
    }
    }

#ifdef AGG_DEBUG
    for (i = 0; i < nprocs; i++) {
    if (count_my_req_per_proc[i] > 0) {
        FPRINTF(stdout, "data needed from %d (count = %d):\n",
                    i, my_req[i].count);
        for (l = 0; l < my_req[i].count; l++) {
        FPRINTF(stdout, "   off[%d] = %lld, len[%d] = %d\n",
                    l, my_req[i].offsets[l], l, my_req[i].lens[l]);
        }
    }
    }
#endif

    *count_my_req_procs_ptr = count_my_req_procs;
    *buf_idx_ptr = buf_idx;
}

int ADIOI_LUSTRE_Docollect(ADIO_File fd, int contig_access_count,
               ADIO_Offset *len_list, int nprocs)
{
    /* If the processes are non-interleaved, we will check the req_size.
     *   if (avg_req_size > big_req_size) {
     *       docollect = 0;
     *   }
     */

    int i, docollect = 1, big_req_size = 0;
    ADIO_Offset req_size = 0, total_req_size;
    int avg_req_size, total_access_count;

    /* calculate total_req_size and total_access_count */
    for (i = 0; i < contig_access_count; i++)
        req_size += len_list[i];
    MPI_Allreduce(&req_size, &total_req_size, 1, MPI_LONG_LONG_INT, MPI_SUM,
               fd->comm);
    MPI_Allreduce(&contig_access_count, &total_access_count, 1, MPI_INT, MPI_SUM,
               fd->comm);
    /* estimate average req_size */
    avg_req_size = (int)(total_req_size / total_access_count);
    /* get hint of big_req_size */
    big_req_size = fd->hints->fs_hints.lustre.coll_threshold;
    /* Don't perform collective I/O if there are big requests */
    if ((big_req_size > 0) && (avg_req_size > big_req_size))
        docollect = 0;

    return docollect;
}

---