/* this moves all files of the specified dataset in the cache to
* make them accessible to new rank mapping */
static int scr_distribute_files(scr_filemap* map, const scr_reddesc* red, int id)
{
int i, round;
int rc = SCR_SUCCESS;
/* TODO: mark dataset as being distributed in filemap,
* because if we fail in the middle of a distribute,
* we can't trust the contents of the files anymore,
* at which point it should be deleted */
/* clean out any incomplete files before we start */
scr_cache_clean(map);
/* for this dataset, get list of ranks we have data for */
int nranks = 0;
int* ranks = NULL;
scr_filemap_list_ranks_by_dataset(map, id, &nranks, &ranks);
/* walk backwards through the list of ranks, and set our start index
* to the rank which is the first rank that is equal to or higher
* than our own rank -- when we assign round ids below, this offsetting
* helps distribute the load */
int start_index = 0;
int invalid_rank_found = 0;
for (i = nranks-1; i >= 0; i--) {
int rank = ranks[i];
/* pick the first rank whose rank id is equal to or higher than our own */
if (rank >= scr_my_rank_world) {
start_index = i;
}
/* while we're at it, check that the rank is within range */
if (rank < 0 || rank >= scr_ranks_world) {
scr_err("Invalid rank id %d in world of %d @ %s:%d",
rank, scr_ranks_world, __FILE__, __LINE__
);
invalid_rank_found = 1;
}
}
/* check that we didn't find an invalid rank on any process */
if (! scr_alltrue(invalid_rank_found == 0)) {
scr_free(&ranks);
return SCR_FAILURE;
}
/* allocate array to record the rank we can send to in each round */
int* have_rank_by_round = (int*) SCR_MALLOC(sizeof(int) * nranks);
int* send_flag_by_round = (int*) SCR_MALLOC(sizeof(int) * nranks);
/* check that we have all of the files for each rank,
* and determine the round we can send them */
scr_hash* send_hash = scr_hash_new();
scr_hash* recv_hash = scr_hash_new();
for (round = 0; round < nranks; round++) {
/* get the rank id */
int index = (start_index + round) % nranks;
int rank = ranks[index];
/* record the rank indexed by the round number */
have_rank_by_round[round] = rank;
/* assume we won't be sending to this rank in this round */
send_flag_by_round[round] = 0;
/* if we have files for this rank, specify the round we can
* send those files in */
if (scr_bool_have_files(map, id, rank)) {
scr_hash_setf(send_hash, NULL, "%d %d", rank, round);
}
}
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* search for the minimum round we can get our files */
int retrieve_rank = -1;
int retrieve_round = -1;
scr_hash_elem* elem = NULL;
for (elem = scr_hash_elem_first(recv_hash);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the rank id */
int rank = scr_hash_elem_key_int(elem);
/* get the round id */
scr_hash* round_hash = scr_hash_elem_hash(elem);
scr_hash_elem* round_elem = scr_hash_elem_first(round_hash);
char* round_str = scr_hash_elem_key(round_elem);
int round = atoi(round_str);
/* record this round and rank number if it's less than the current round */
if (round < retrieve_round || retrieve_round == -1) {
retrieve_round = round;
retrieve_rank = rank;
}
}
/* done with the round hashes, free them off */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
/* free off our list of ranks */
scr_free(&ranks);
/* for some redundancy schemes, we know at this point whether we
* can recover all files */
int can_get_files = (retrieve_rank != -1);
if (red->copy_type != SCR_COPY_XOR && !scr_alltrue(can_get_files)) {
/* print a debug message indicating which rank is missing files */
if (! can_get_files) {
scr_dbg(2, "Cannot find process that has my checkpoint files @ %s:%d",
__FILE__, __LINE__
);
}
return SCR_FAILURE;
}
/* get the maximum retrieve round */
int max_rounds = 0;
MPI_Allreduce(
&retrieve_round, &max_rounds, 1, MPI_INT, MPI_MAX, scr_comm_world
);
/* tell destination which round we'll take our files in */
send_hash = scr_hash_new();
recv_hash = scr_hash_new();
if (retrieve_rank != -1) {
scr_hash_setf(send_hash, NULL, "%d %d", retrieve_rank, retrieve_round);
}
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* determine which ranks want to fetch their files from us */
for(elem = scr_hash_elem_first(recv_hash);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the round id */
scr_hash* round_hash = scr_hash_elem_hash(elem);
scr_hash_elem* round_elem = scr_hash_elem_first(round_hash);
char* round_str = scr_hash_elem_key(round_elem);
int round = atoi(round_str);
/* record whether this rank wants its files from us */
if (round >= 0 && round < nranks) {
send_flag_by_round[round] = 1;
}
}
/* done with the round hashes, free them off */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
int tmp_rc = 0;
/* run through rounds and exchange files */
for (round = 0; round <= max_rounds; round++) {
/* assume we don't need to send or receive any files this round */
int send_rank = MPI_PROC_NULL;
int recv_rank = MPI_PROC_NULL;
int send_num = 0;
int recv_num = 0;
/* check whether I can potentially send to anyone in this round */
if (round < nranks) {
/* have someone's files, check whether they are asking
* for them this round */
if (send_flag_by_round[round]) {
/* need to send files this round, remember to whom and how many */
int dst_rank = have_rank_by_round[round];
send_rank = dst_rank;
send_num = scr_filemap_num_files(map, id, dst_rank);
}
}
/* if I'm supposed to get my files this round, set the recv_rank */
if (retrieve_round == round) {
recv_rank = retrieve_rank;
}
/* TODO: another special case is to just move files if the
* processes are on the same node */
/* if i'm sending to myself, just move (rename) each file */
if (send_rank == scr_my_rank_world) {
/* get our file list */
int numfiles = 0;
char** files = NULL;
scr_filemap_list_files(map, id, send_rank, &numfiles, &files);
/* TODO: sort files in reverse order by size */
/* iterate over and rename each file */
for (i=0; i < numfiles; i++) {
/* get the current file name */
char* file = files[i];
/* lookup meta data for this file */
scr_meta* meta = scr_meta_new();
scr_filemap_get_meta(map, id, send_rank, file, meta);
/* get the path for this file based on its type
* and dataset id */
char* dir = NULL;
if (scr_meta_check_filetype(meta, SCR_META_FILE_USER) == SCR_SUCCESS) {
dir = scr_cache_dir_get(red, id);
} else {
dir = scr_cache_dir_hidden_get(red, id);
}
/* build the new file name */
scr_path* path_newfile = scr_path_from_str(file);
scr_path_basename(path_newfile);
scr_path_prepend_str(path_newfile, dir);
char* newfile = scr_path_strdup(path_newfile);
/* if the new file name is different from the old name, rename it */
if (strcmp(file, newfile) != 0) {
/* record the new filename to our map and write it to disk */
scr_filemap_add_file(map, id, send_rank, newfile);
scr_filemap_set_meta(map, id, send_rank, newfile, meta);
scr_filemap_write(scr_map_file, map);
/* rename the file */
scr_dbg(2, "Round %d: rename(%s, %s)", round, file, newfile);
tmp_rc = rename(file, newfile);
if (tmp_rc != 0) {
/* TODO: to cross mount points, if tmp_rc == EXDEV,
* open new file, copy, and delete orig */
scr_err("Moving checkpoint file: rename(%s, %s) %s errno=%d @ %s:%d",
file, newfile, strerror(errno), errno, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* remove the old name from the filemap and write it to disk */
scr_filemap_remove_file(map, id, send_rank, file);
scr_filemap_write(scr_map_file, map);
}
/* free the path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
/* free directory string */
scr_free(&dir);
/* free meta data */
scr_meta_delete(&meta);
}
/* free the list of filename pointers */
scr_free(&files);
} else {
/* if we have files for this round, but the correspdonding
* rank doesn't need them, delete the files */
if (round < nranks && send_rank == MPI_PROC_NULL) {
int dst_rank = have_rank_by_round[round];
scr_unlink_rank(map, id, dst_rank);
}
/* sending to and/or recieving from another node */
if (send_rank != MPI_PROC_NULL || recv_rank != MPI_PROC_NULL) {
/* have someone to send to or receive from */
int have_outgoing = 0;
int have_incoming = 0;
if (send_rank != MPI_PROC_NULL) {
have_outgoing = 1;
}
if (recv_rank != MPI_PROC_NULL) {
have_incoming = 1;
}
/* first, determine how many files I will be receiving and
* tell how many I will be sending */
MPI_Request request[2];
MPI_Status status[2];
int num_req = 0;
if (have_incoming) {
MPI_Irecv(
&recv_num, 1, MPI_INT, recv_rank, 0,
scr_comm_world, &request[num_req]
);
num_req++;
}
if (have_outgoing) {
MPI_Isend(
&send_num, 1, MPI_INT, send_rank, 0,
scr_comm_world, &request[num_req]
);
num_req++;
}
if (num_req > 0) {
MPI_Waitall(num_req, request, status);
}
/* record how many files I will receive (need to distinguish
* between 0 files and not knowing) */
if (have_incoming) {
scr_filemap_set_expected_files(map, id, scr_my_rank_world, recv_num);
}
/* turn off send or receive flags if the file count is 0,
* nothing else to do */
if (send_num == 0) {
have_outgoing = 0;
send_rank = MPI_PROC_NULL;
}
if (recv_num == 0) {
have_incoming = 0;
recv_rank = MPI_PROC_NULL;
}
/* TODO: since we overwrite files in place in order to avoid
* running out of storage space, we should sort files in order
* of descending size for the next step */
/* get our file list for the destination */
int numfiles = 0;
char** files = NULL;
if (have_outgoing) {
scr_filemap_list_files(map, id, send_rank, &numfiles, &files);
}
/* while we have a file to send or receive ... */
while (have_incoming || have_outgoing) {
/* get the filename */
char* file = NULL;
scr_meta* send_meta = NULL;
if (have_outgoing) {
file = files[numfiles - send_num];
send_meta = scr_meta_new();
scr_filemap_get_meta(map, id, send_rank, file, send_meta);
}
/* exchange meta data so we can determine type of incoming file */
scr_meta* recv_meta = scr_meta_new();
scr_hash_sendrecv(send_meta, send_rank, recv_meta, recv_rank, scr_comm_world);
/* get the path for this file based on its type and dataset id */
char* dir = NULL;
if (have_incoming) {
if (scr_meta_check_filetype(recv_meta, SCR_META_FILE_USER) == SCR_SUCCESS) {
dir = scr_cache_dir_get(red, id);
} else {
dir = scr_cache_dir_hidden_get(red, id);
}
}
/* exhange file names with partners,
* building full path of incoming file */
char file_partner[SCR_MAX_FILENAME];
scr_swap_file_names(
file, send_rank, file_partner, sizeof(file_partner), recv_rank,
dir, scr_comm_world
);
/* free directory string */
scr_free(&dir);
/* free incoming meta data (we'll get this again later) */
scr_meta_delete(&recv_meta);
/* if we'll receive a file, record the name of our file
* in the filemap and write it to disk */
recv_meta = NULL;
if (recv_rank != MPI_PROC_NULL) {
recv_meta = scr_meta_new();
scr_filemap_add_file(map, id, scr_my_rank_world, file_partner);
scr_filemap_write(scr_map_file, map);
}
/* either sending or receiving a file this round, since we move files,
* it will be deleted or overwritten */
if (scr_swap_files(MOVE_FILES, file, send_meta, send_rank,
file_partner, recv_meta, recv_rank, scr_comm_world) != SCR_SUCCESS)
{
scr_err("Swapping files: %s to %d, %s from %d @ %s:%d",
file, send_rank, file_partner, recv_rank, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* if we received a file, record its meta data and decrement
* our receive count */
if (have_incoming) {
/* record meta data for the file we received */
scr_filemap_set_meta(map, id, scr_my_rank_world, file_partner, recv_meta);
scr_meta_delete(&recv_meta);
/* decrement receive count */
recv_num--;
if (recv_num == 0) {
have_incoming = 0;
recv_rank = MPI_PROC_NULL;
}
}
/* if we sent a file, remove it from the filemap and decrement
* our send count */
if (have_outgoing) {
/* remove file from the filemap */
scr_filemap_remove_file(map, id, send_rank, file);
scr_meta_delete(&send_meta);
/* decrement our send count */
send_num--;
if (send_num == 0) {
have_outgoing = 0;
send_rank = MPI_PROC_NULL;
}
}
/* update filemap on disk */
scr_filemap_write(scr_map_file, map);
}
/* free our file list */
scr_free(&files);
}
}
}
/* if we have more rounds than max rounds, delete the remainder of our files */
for (round = max_rounds+1; round < nranks; round++) {
/* have someone's files for this round, so delete them */
int dst_rank = have_rank_by_round[round];
scr_unlink_rank(map, id, dst_rank);
}
scr_free(&send_flag_by_round);
scr_free(&have_rank_by_round);
/* write out new filemap and free the memory resources */
scr_filemap_write(scr_map_file, map);
/* clean out any incomplete files */
scr_cache_clean(map);
/* TODO: if the exchange or redundancy rebuild failed,
* we should also delete any *good* files we received */
/* return whether distribute succeeded, it does not ensure we have
* all of our files, only that the transfer completed without failure */
return rc;
}
/* apply redundancy scheme to file and return number of bytes copied
* in bytes parameter */
int scr_reddesc_apply(
scr_filemap* map,
const scr_reddesc* c,
int id,
double* bytes)
{
/* initialize to 0 */
*bytes = 0.0;
/* step through each of my files for the specified dataset
* to scan for any incomplete files */
int valid = 1;
double my_bytes = 0.0;
scr_hash_elem* file_elem;
for (file_elem = scr_filemap_first_file(map, id, scr_my_rank_world);
file_elem != NULL;
file_elem = scr_hash_elem_next(file_elem))
{
/* get the filename */
char* file = scr_hash_elem_key(file_elem);
/* check the file */
if (! scr_bool_have_file(map, id, scr_my_rank_world, file, scr_ranks_world)) {
scr_dbg(2, "File determined to be invalid: %s", file);
valid = 0;
}
/* add up the number of bytes on our way through */
my_bytes += (double) scr_file_size(file);
/* if crc_on_copy is set, compute crc and update meta file
* (PARTNER does this during the copy) */
if (scr_crc_on_copy && c->copy_type != SCR_COPY_PARTNER) {
scr_compute_crc(map, id, scr_my_rank_world, file);
}
}
/* determine whether everyone's files are good */
int all_valid = scr_alltrue(valid);
if (! all_valid) {
if (scr_my_rank_world == 0) {
scr_dbg(1, "Exiting copy since one or more checkpoint files is invalid");
}
return SCR_FAILURE;
}
/* start timer */
time_t timestamp_start;
double time_start;
if (scr_my_rank_world == 0) {
timestamp_start = scr_log_seconds();
time_start = MPI_Wtime();
}
/* apply the redundancy scheme */
int rc = SCR_FAILURE;
switch (c->copy_type) {
case SCR_COPY_SINGLE:
rc = SCR_SUCCESS;
break;
case SCR_COPY_PARTNER:
rc = scr_reddesc_apply_partner(map, c, id);
break;
case SCR_COPY_XOR:
rc = scr_reddesc_apply_xor(map, c, id);
break;
}
/* record the number of files this task wrote during this dataset
* (need to remember when a task writes 0 files) */
int num_files = scr_filemap_num_files(map, id, scr_my_rank_world);
scr_filemap_set_expected_files(map, id, scr_my_rank_world, num_files);
scr_filemap_write(scr_map_file, map);
/* determine whether everyone succeeded in their copy */
int valid_copy = (rc == SCR_SUCCESS);
if (! valid_copy) {
scr_err("scr_copy_files failed with return code %d @ %s:%d",
rc, __FILE__, __LINE__
);
}
int all_valid_copy = scr_alltrue(valid_copy);
rc = all_valid_copy ? SCR_SUCCESS : SCR_FAILURE;
/* add up total number of bytes */
MPI_Allreduce(&my_bytes, bytes, 1, MPI_DOUBLE, MPI_SUM, scr_comm_world);
/* stop timer and report performance info */
if (scr_my_rank_world == 0) {
double time_end = MPI_Wtime();
double time_diff = time_end - time_start;
double bw = *bytes / (1024.0 * 1024.0 * time_diff);
scr_dbg(1, "scr_reddesc_apply: %f secs, %e bytes, %f MB/s, %f MB/s per proc",
time_diff, *bytes, bw, bw/scr_ranks_world
);
/* log data on the copy in the database */
if (scr_log_enable) {
char* dir = scr_cache_dir_get(c, id);
scr_log_transfer("COPY", c->base, dir, &id, ×tamp_start, &time_diff, bytes);
scr_free(&dir);
}
}
return rc;
}
/* fetch files from parallel file system */
static int scr_fetch_files(
scr_filemap* map,
scr_path* fetch_path,
int* dataset_id,
int* checkpoint_id)
{
/* get fetch directory as string */
char* fetch_dir = scr_path_strdup(fetch_path);
/* this may take a while, so tell user what we're doing */
if (scr_my_rank_world == 0) {
scr_dbg(1, "Attempting fetch from %s", fetch_dir);
}
/* make sure all processes make it this far before progressing */
MPI_Barrier(scr_comm_world);
/* start timer */
time_t timestamp_start;
double time_start;
if (scr_my_rank_world == 0) {
timestamp_start = scr_log_seconds();
time_start = MPI_Wtime();
}
/* log the fetch attempt */
if (scr_my_rank_world == 0) {
if (scr_log_enable) {
time_t now = scr_log_seconds();
scr_log_event("FETCH STARTED", fetch_dir, NULL, &now, NULL);
}
}
/* allocate a new hash to get a list of files to fetch */
scr_hash* file_list = scr_hash_new();
/* read the summary file */
if (scr_fetch_summary(fetch_dir, file_list) != SCR_SUCCESS) {
if (scr_my_rank_world == 0) {
scr_dbg(1, "Failed to read summary file @ %s:%d", __FILE__, __LINE__);
if (scr_log_enable) {
double time_end = MPI_Wtime();
double time_diff = time_end - time_start;
time_t now = scr_log_seconds();
scr_log_event("FETCH FAILED", fetch_dir, NULL, &now, &time_diff);
}
}
scr_hash_delete(&file_list);
scr_free(&fetch_dir);
return SCR_FAILURE;
}
/* get a pointer to the dataset */
scr_dataset* dataset = scr_hash_get(file_list, SCR_KEY_DATASET);
/* get the dataset id */
int id;
if (scr_dataset_get_id(dataset, &id) != SCR_SUCCESS) {
if (scr_my_rank_world == 0) {
scr_dbg(1, "Invalid id in summary file @ %s:%d", __FILE__, __LINE__);
if (scr_log_enable) {
double time_end = MPI_Wtime();
double time_diff = time_end - time_start;
time_t now = scr_log_seconds();
scr_log_event("FETCH FAILED", fetch_dir, NULL, &now, &time_diff);
}
}
scr_hash_delete(&file_list);
scr_free(&fetch_dir);
return SCR_FAILURE;
}
/* get the checkpoint id for this dataset */
int ckpt_id;
if (scr_dataset_get_ckpt(dataset, &ckpt_id) != SCR_SUCCESS) {
/* eventually, we'll support reading of non-checkpoint datasets,
* but we don't yet */
scr_err("Failed to read checkpoint id from dataset @ %s:%d",
__FILE__, __LINE__
);
scr_hash_delete(&file_list);
scr_free(&fetch_dir);
return SCR_FAILURE;
}
/* delete any existing files for this dataset id (do this before
* filemap_read) */
scr_cache_delete(map, id);
/* get the redundancy descriptor for this id */
scr_reddesc* c = scr_reddesc_for_checkpoint(ckpt_id, scr_nreddescs, scr_reddescs);
/* store our redundancy descriptor hash in the filemap */
scr_hash* my_desc_hash = scr_hash_new();
scr_reddesc_store_to_hash(c, my_desc_hash);
scr_filemap_set_desc(map, id, scr_my_rank_world, my_desc_hash);
scr_hash_delete(&my_desc_hash);
/* write the filemap out before creating the directory */
scr_filemap_write(scr_map_file, map);
/* create the cache directory */
scr_cache_dir_create(c, id);
/* get the cache directory */
char cache_dir[SCR_MAX_FILENAME];
scr_cache_dir_get(c, id, cache_dir);
/* now we can finally fetch the actual files */
int success = 1;
if (scr_fetch_data(file_list, cache_dir, map) != SCR_SUCCESS) {
success = 0;
}
/* free the hash holding the summary file data */
scr_hash_delete(&file_list);
/* check that all processes copied their file successfully */
if (! scr_alltrue(success)) {
/* someone failed, so let's delete the partial checkpoint */
scr_cache_delete(map, id);
if (scr_my_rank_world == 0) {
scr_dbg(1, "One or more processes failed to read its files @ %s:%d",
__FILE__, __LINE__
);
if (scr_log_enable) {
double time_end = MPI_Wtime();
double time_diff = time_end - time_start;
time_t now = scr_log_seconds();
scr_log_event("FETCH FAILED", fetch_dir, &id, &now, &time_diff);
}
}
scr_free(&fetch_dir);
return SCR_FAILURE;
}
/* apply redundancy scheme */
double bytes_copied = 0.0;
int rc = scr_reddesc_apply(map, c, id, &bytes_copied);
if (rc == SCR_SUCCESS) {
/* record dataset and checkpoint ids */
*dataset_id = id;
*checkpoint_id = ckpt_id;
/* update our flush file to indicate this checkpoint is in cache
* as well as the parallel file system */
/* TODO: should we place SCR_FLUSH_KEY_LOCATION_PFS before
* scr_reddesc_apply? */
scr_flush_file_location_set(id, SCR_FLUSH_KEY_LOCATION_CACHE);
scr_flush_file_location_set(id, SCR_FLUSH_KEY_LOCATION_PFS);
scr_flush_file_location_unset(id, SCR_FLUSH_KEY_LOCATION_FLUSHING);
} else {
/* something went wrong, so delete this checkpoint from the cache */
scr_cache_delete(scr_map, id);
}
/* stop timer, compute bandwidth, and report performance */
double total_bytes = bytes_copied;
if (scr_my_rank_world == 0) {
double time_end = MPI_Wtime();
double time_diff = time_end - time_start;
double bw = total_bytes / (1024.0 * 1024.0 * time_diff);
scr_dbg(1, "scr_fetch_files: %f secs, %e bytes, %f MB/s, %f MB/s per proc",
time_diff, total_bytes, bw, bw/scr_ranks_world
);
/* log data on the fetch to the database */
if (scr_log_enable) {
time_t now = scr_log_seconds();
if (rc == SCR_SUCCESS) {
scr_log_event("FETCH SUCCEEDED", fetch_dir, &id, &now, &time_diff);
} else {
scr_log_event("FETCH FAILED", fetch_dir, &id, &now, &time_diff);
}
char cache_dir[SCR_MAX_FILENAME];
scr_cache_dir_get(c, id, cache_dir);
scr_log_transfer("FETCH", fetch_dir, cache_dir, &id,
×tamp_start, &time_diff, &total_bytes
);
}
}
/* free fetch direcotry string */
scr_free(&fetch_dir);
return rc;
}
/* copy files to a partner node */
static int scr_reddesc_apply_partner(
scr_filemap* map,
const scr_reddesc* c,
int id)
{
int rc = SCR_SUCCESS;
/* get pointer to partner state structure */
scr_reddesc_partner* state = (scr_reddesc_partner*) c->copy_state;
/* get a list of our files */
int numfiles = 0;
char** files = NULL;
scr_filemap_list_files(map, id, scr_my_rank_world, &numfiles, &files);
/* first, determine how many files we'll be sending and receiving
* with our partners */
MPI_Status status;
int send_num = numfiles;
int recv_num = 0;
MPI_Sendrecv(
&send_num, 1, MPI_INT, state->rhs_rank, 0,
&recv_num, 1, MPI_INT, state->lhs_rank, 0,
c->comm, &status
);
/* record how many files our partner will send */
scr_filemap_set_expected_files(map, id, state->lhs_rank_world, recv_num);
/* remember which node our partner is on (needed for scavenge) */
scr_hash* flushdesc = scr_hash_new();
scr_filemap_get_flushdesc(map, id, state->lhs_rank_world, flushdesc);
scr_hash_util_set_int(flushdesc, SCR_SCAVENGE_KEY_PRESERVE, scr_preserve_directories);
scr_hash_util_set_int(flushdesc, SCR_SCAVENGE_KEY_CONTAINER, scr_use_containers);
scr_hash_util_set_str(flushdesc, SCR_SCAVENGE_KEY_PARTNER, state->lhs_hostname);
scr_filemap_set_flushdesc(map, id, state->lhs_rank_world, flushdesc);
scr_hash_delete(&flushdesc);
/* record partner's redundancy descriptor hash */
scr_hash* lhs_desc_hash = scr_hash_new();
scr_hash* my_desc_hash = scr_hash_new();
scr_reddesc_store_to_hash(c, my_desc_hash);
scr_hash_sendrecv(my_desc_hash, state->rhs_rank, lhs_desc_hash, state->lhs_rank, c->comm);
scr_filemap_set_desc(map, id, state->lhs_rank_world, lhs_desc_hash);
scr_hash_delete(&my_desc_hash);
scr_hash_delete(&lhs_desc_hash);
/* store this info in our filemap before we receive any files */
scr_filemap_write(scr_map_file, map);
/* define directory to receive partner file in */
char* dir = scr_cache_dir_get(c, id);
/* for each potential file, step through a call to swap */
while (send_num > 0 || recv_num > 0) {
/* assume we won't send or receive in this step */
int send_rank = MPI_PROC_NULL;
int recv_rank = MPI_PROC_NULL;
/* if we have a file left to send,
* get the filename and destination rank */
char* file = NULL;
if (send_num > 0) {
int i = numfiles - send_num;
file = files[i];
send_rank = state->rhs_rank;
send_num--;
}
/* if we have a file left to receive, get the rank */
if (recv_num > 0) {
recv_rank = state->lhs_rank;
recv_num--;
}
/* exhange file names with partners */
char file_partner[SCR_MAX_FILENAME];
scr_swap_file_names(file, send_rank, file_partner, sizeof(file_partner), recv_rank, dir, c->comm);
/* if we'll receive a file, record the name of our partner's
* file in the filemap */
if (recv_rank != MPI_PROC_NULL) {
scr_filemap_add_file(map, id, state->lhs_rank_world, file_partner);
scr_filemap_write(scr_map_file, map);
}
/* get meta data of file we're sending */
scr_meta* send_meta = scr_meta_new();
scr_filemap_get_meta(map, id, scr_my_rank_world, file, send_meta);
/* exhange files with partners */
scr_meta* recv_meta = scr_meta_new();
if (scr_swap_files(COPY_FILES, file, send_meta, send_rank, file_partner, recv_meta, recv_rank, c->comm) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
scr_filemap_set_meta(map, id, state->lhs_rank_world, file_partner, recv_meta);
/* free meta data for these files */
scr_meta_delete(&recv_meta);
scr_meta_delete(&send_meta);
}
/* free cache directory string */
scr_free(&dir);
/* write out the updated filemap */
scr_filemap_write(scr_map_file, map);
/* free our list of files */
scr_free(&files);
return rc;
}
