/* allocates a new hash element */
static scr_hash_elem* scr_hash_elem_new()
{
scr_hash_elem* elem = (scr_hash_elem*) SCR_MALLOC(sizeof(scr_hash_elem));
elem->key = NULL;
elem->hash = NULL;
return elem;
}
/* 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;
}
/* since on a restart we may end up with more or fewer ranks on a node than the
* previous run, rely on the master to read in and distribute the filemap to
* other ranks on the node */
int scr_scatter_filemaps(scr_filemap* my_map)
{
/* TODO: if the control directory is on a device shared by lots of procs,
* we should read and distribute this data in a more scalable way */
/* allocate empty send hash */
scr_hash* send_hash = scr_hash_new();
/* if i'm the master on this node, read in all filemaps */
if (scr_storedesc_cntl->rank == 0) {
/* create an empty filemap */
scr_filemap* all_map = scr_filemap_new();
/* read in the master map */
scr_hash* hash = scr_hash_new();
scr_hash_read_path(scr_master_map_file, hash);
/* for each filemap listed in the master map */
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(scr_hash_get(hash, "Filemap"));
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the filename of this filemap */
char* file = scr_hash_elem_key(elem);
/* TODO MEMFS: mount storage for each filemap */
/* read in the filemap */
scr_filemap* tmp_map = scr_filemap_new();
scr_path* path_file = scr_path_from_str(file);
scr_filemap_read(path_file, tmp_map);
scr_path_delete(&path_file);
/* merge it with the all_map */
scr_filemap_merge(all_map, tmp_map);
/* delete filemap */
scr_filemap_delete(&tmp_map);
/* TODO: note that if we fail after unlinking this file but before
* writing out the new file, we'll lose information */
/* delete the file */
scr_file_unlink(file);
}
/* free the hash object */
scr_hash_delete(&hash);
/* write out new local 0 filemap */
if (scr_filemap_num_ranks(all_map) > 0) {
scr_filemap_write(scr_map_file, all_map);
}
/* get global rank of each rank */
int* ranks = (int*) SCR_MALLOC(scr_storedesc_cntl->ranks * sizeof(int));
MPI_Gather(
&scr_my_rank_world, 1, MPI_INT, ranks, 1, MPI_INT,
0, scr_storedesc_cntl->comm
);
/* for each rank, send them their own file data if we have it */
int i;
for (i=0; i < scr_storedesc_cntl->ranks; i++) {
int rank = ranks[i];
if (scr_filemap_have_rank(all_map, rank)) {
/* extract the filemap for this rank */
scr_filemap* tmp_map = scr_filemap_extract_rank(all_map, rank);
/* get a reference to the hash object that we'll send to this rank,
* and merge this filemap into it */
scr_hash* tmp_hash = scr_hash_getf(send_hash, "%d", i);
if (tmp_hash == NULL) {
/* if we don't find an existing entry in the send_hash,
* create an empty hash and insert it */
scr_hash* empty_hash = scr_hash_new();
scr_hash_setf(send_hash, empty_hash, "%d", i);
tmp_hash = empty_hash;
}
scr_hash_merge(tmp_hash, tmp_map);
/* delete the filemap for this rank */
scr_filemap_delete(&tmp_map);
}
}
/* free our rank list */
scr_free(&ranks);
/* now just round robin the remainder across the set (load balancing) */
int num;
int* remaining_ranks = NULL;
scr_filemap_list_ranks(all_map, &num, &remaining_ranks);
int j = 0;
while (j < num) {
/* pick a rank in to send to */
i = j % scr_storedesc_cntl->ranks;
/* extract the filemap for this rank */
scr_filemap* tmp_map = scr_filemap_extract_rank(all_map, remaining_ranks[j]);
/* get a reference to the hash object that we'll send to this rank,
* and merge this filemap into it */
scr_hash* tmp_hash = scr_hash_getf(send_hash, "%d", i);
if (tmp_hash == NULL) {
/* if we don't find an existing entry in the send_hash,
* create an empty hash and insert it */
scr_hash* empty_hash = scr_hash_new();
scr_hash_setf(send_hash, empty_hash, "%d", i);
tmp_hash = empty_hash;
}
scr_hash_merge(tmp_hash, tmp_map);
/* delete the filemap for this rank */
scr_filemap_delete(&tmp_map);
j++;
}
scr_free(&remaining_ranks);
/* delete the filemap */
scr_filemap_delete(&all_map);
/* write out the new master filemap */
hash = scr_hash_new();
char file[SCR_MAX_FILENAME];
for (i=0; i < scr_storedesc_cntl->ranks; i++) {
sprintf(file, "%s/filemap_%d.scrinfo", scr_cntl_prefix, i);
scr_hash_set_kv(hash, "Filemap", file);
}
scr_hash_write_path(scr_master_map_file, hash);
scr_hash_delete(&hash);
} else {
/* send our global rank to the master */
MPI_Gather(
&scr_my_rank_world, 1, MPI_INT, NULL, 1, MPI_INT,
0, scr_storedesc_cntl->comm
);
}
/* receive our filemap from master */
scr_hash* recv_hash = scr_hash_new();
scr_hash_exchange(send_hash, recv_hash, scr_storedesc_cntl->comm);
/* merge map sent from master into our map */
scr_hash* map_from_master = scr_hash_getf(recv_hash, "%d", 0);
if (map_from_master != NULL) {
scr_hash_merge(my_map, map_from_master);
}
/* write out our local filemap */
if (scr_filemap_num_ranks(my_map) > 0) {
scr_filemap_write(scr_map_file, my_map);
}
/* free off our send and receive hashes */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
return SCR_SUCCESS;
}
/* given a filename, its meta data, its list of segments, and list of destination containers,
* copy file to container files */
static int scr_flush_file_to_containers(
const char* file,
scr_meta* meta,
scr_hash* segments,
const char* dst_dir)
{
/* check that we got something for a source file */
if (file == NULL || strcmp(file, "") == 0) {
scr_err("Invalid source file @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* check that our other arguments are valid */
if (meta == NULL || segments == NULL) {
scr_err("Invalid metadata or segments @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* open the file for reading */
int fd_src = scr_open(file, O_RDONLY);
if (fd_src < 0) {
scr_err("Opening file to copy: scr_open(%s) errno=%d %s @ %s:%d",
file, errno, strerror(errno), __FILE__, __LINE__
);
return SCR_FAILURE;
}
#if !defined(__APPLE__)
/* TODO:
posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL)
that tells the kernel that you don't ever need the pages
from the file again, and it won't bother keeping them in the page cache.
*/
posix_fadvise(fd_src, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL);
#endif
/* get the buffer size we'll use to write to the file */
unsigned long buf_size = scr_file_buf_size;
/* allocate buffer to read in file chunks */
char* buf = (char*) SCR_MALLOC(buf_size);
/* initialize crc value */
uLong crc;
if (scr_crc_on_flush) {
crc = crc32(0L, Z_NULL, 0);
}
int rc = SCR_SUCCESS;
/* write out each segment */
scr_hash_sort_int(segments, SCR_HASH_SORT_ASCENDING);
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(segments);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the container info for this segment */
scr_hash* hash = scr_hash_elem_hash(elem);
/* get the offset into the container and the length of the segment (both in bytes) */
char* container_name;
unsigned long container_offset, segment_length;
if (scr_container_get_name_offset_length(hash,
&container_name, &container_offset, &segment_length) != SCR_SUCCESS)
{
scr_err("Failed to get segment offset and length @ %s:%d",
__FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
/* build full name to destination file */
scr_path* dst_path = scr_path_from_str(dst_dir);
scr_path_append_str(dst_path, container_name);
scr_path_reduce(dst_path);
char* dst_file = scr_path_strdup(dst_path);
/* open container file for writing -- we don't truncate here because more than one
* process may be writing to the same file */
int fd_container = scr_open(dst_file, O_WRONLY);
if (fd_container < 0) {
scr_err("Opening file for writing: scr_open(%s) errno=%d %s @ %s:%d",
dst_file, errno, strerror(errno), __FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
#if !defined(__APPLE__)
/* TODO:
posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL)
that tells the kernel that you don't ever need the pages
from the file again, and it won't bother keeping them in the page cache.
*/
posix_fadvise(fd_container, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL);
#endif
/* seek to offset within container */
off_t pos = (off_t) container_offset;
if (lseek(fd_container, pos, SEEK_SET) == (off_t)-1) {
/* our seek failed, return an error */
scr_err("Failed to seek to byte %lu in %s @ %s:%d",
pos, dst_file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
/* copy data from file into container in chunks */
unsigned long remaining = segment_length;
while (remaining > 0) {
/* read / write up to buf_size bytes at a time from file */
unsigned long count = remaining;
if (count > buf_size) {
count = buf_size;
}
/* attempt to read buf_size bytes from file */
int nread = scr_read_attempt(file, fd_src, buf, count);
/* if we read some bytes, write them out */
if (nread > 0) {
/* optionally compute crc value as we go */
if (scr_crc_on_flush) {
crc = crc32(crc, (const Bytef*) buf, (uInt) nread);
}
/* write our nread bytes out */
int nwrite = scr_write_attempt(dst_file, fd_container, buf, nread);
/* check for a write error or a short write */
if (nwrite != nread) {
/* write had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
/* subtract the bytes we've processed from the number remaining */
remaining -= (unsigned long) nread;
}
/* assume a short read is an error */
if (nread < count) {
/* read had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
/* check for a read error, stop copying and return an error */
if (nread < 0) {
/* read had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
}
/* close container */
if (scr_close(dst_file, fd_container) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* free the container file name and path */
scr_free(&dst_file);
scr_path_delete(&dst_path);
}
/* close the source file */
if (scr_close(file, fd_src) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* free buffer */
scr_free(&buf);
/* verify / set crc value */
if (rc == SCR_SUCCESS) {
uLong crc2;
if (scr_crc_on_flush) {
if (scr_meta_get_crc32(meta, &crc2) == SCR_SUCCESS) {
/* if a crc is already set in the meta data, check that we computed the same value */
if (crc != crc2) {
scr_err("CRC32 mismatch detected when flushing file %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
} else {
/* if there is no crc set, let's set it now */
scr_meta_set_crc32(meta, crc);
}
}
}
return rc;
}
/* flushes data for files specified in file_list (with flow control),
* and records status of each file in data */
static int scr_flush_data(scr_hash* file_list, scr_hash* data)
{
int flushed = SCR_SUCCESS;
/* flow control the write among processes */
if (scr_my_rank_world == 0) {
/* first, flush each of my files and fill in meta data structure */
if (scr_flush_files_list(file_list, data) != SCR_SUCCESS) {
flushed = SCR_FAILURE;
}
/* now, have a sliding window of w processes write simultaneously */
int w = scr_flush_width;
if (w > (scr_ranks_world - 1)) {
w = scr_ranks_world - 1;
}
/* allocate MPI_Request arrays and an array of ints */
int* flags = (int*) SCR_MALLOC(2 * w * sizeof(int));
MPI_Request* req = (MPI_Request*) SCR_MALLOC(2 * w * sizeof(MPI_Request));
MPI_Status status;
int i = 1;
int outstanding = 0;
int index = 0;
while (i < scr_ranks_world || outstanding > 0) {
/* issue up to w outstanding sends and receives */
while (i < scr_ranks_world && outstanding < w) {
/* post a receive for the response message we'll get back when rank i is done */
MPI_Irecv(&flags[w + index], 1, MPI_INT, i, 0, scr_comm_world, &req[w + index]);
/* post a send to tell rank i to start */
flags[index] = flushed;
MPI_Isend(&flags[index], 1, MPI_INT, i, 0, scr_comm_world, &req[index]);
/* update the number of outstanding requests */
i++;
outstanding++;
index++;
}
/* wait to hear back from any rank */
MPI_Waitany(w, &req[w], &index, &status);
/* someone responded, the send to this rank should also be done, so complete it */
MPI_Wait(&req[index], &status);
/* determine whether this rank flushed its file successfully */
if (flags[w + index] != SCR_SUCCESS) {
flushed = SCR_FAILURE;
}
/* one less request outstanding now */
outstanding--;
}
/* free the MPI_Request arrays */
scr_free(&req);
scr_free(&flags);
} else {
/* receive signal to start */
int start = 0;
MPI_Status status;
MPI_Recv(&start, 1, MPI_INT, 0, 0, scr_comm_world, &status);
/* flush files if we've had success so far, otherwise skip the flush and return failure */
if (start == SCR_SUCCESS) {
/* flush each of my files and fill in meta data strucutre */
if (scr_flush_files_list(file_list, data) != SCR_SUCCESS) {
flushed = SCR_FAILURE;
}
} else {
/* someone failed before we even started, so don't bother */
flushed = SCR_FAILURE;
}
/* send message to rank 0 to report that we're done */
MPI_Send(&flushed, 1, MPI_INT, 0, 0, scr_comm_world);
}
/* determine whether everyone wrote their files ok */
if (scr_alltrue((flushed == SCR_SUCCESS))) {
return SCR_SUCCESS;
}
return SCR_FAILURE;
}
/* allocates a new hash */
scr_hash* scr_hash_new()
{
scr_hash* hash = (scr_hash*) SCR_MALLOC(sizeof(scr_hash));
LIST_INIT(hash);
return hash;
}
/* fetch files specified in file_list into specified dir and update
* filemap */
static int scr_fetch_data(
const scr_hash* file_list,
const char* dir,
scr_filemap* map)
{
int success = SCR_SUCCESS;
/* flow control rate of file reads from rank 0 */
if (scr_my_rank_world == 0) {
/* fetch these files into the directory */
if (scr_fetch_files_list(file_list, dir, map) != SCR_SUCCESS) {
success = SCR_FAILURE;
}
/* now, have a sliding window of w processes read simultaneously */
int w = scr_fetch_width;
if (w > scr_ranks_world-1) {
w = scr_ranks_world-1;
}
/* allocate MPI_Request arrays and an array of ints */
int* flags = (int*) SCR_MALLOC(2 * w * sizeof(int));
MPI_Request* req = (MPI_Request*) SCR_MALLOC(2 * w * sizeof(MPI_Request));
MPI_Status status;
/* execute our flow control window */
int outstanding = 0;
int index = 0;
int i = 1;
while (i < scr_ranks_world || outstanding > 0) {
/* issue up to w outstanding sends and receives */
while (i < scr_ranks_world && outstanding < w) {
/* post a receive for the response message we'll get back when
* rank i is done */
MPI_Irecv(&flags[index + w], 1, MPI_INT, i, 0, scr_comm_world, &req[index + w]);
/* send a start signal to this rank */
flags[index] = success;
MPI_Isend(&flags[index], 1, MPI_INT, i, 0, scr_comm_world, &req[index]);
/* update the number of outstanding requests */
outstanding++;
index++;
i++;
}
/* wait to hear back from any rank */
MPI_Waitany(w, &req[w], &index, &status);
/* the corresponding send must be complete */
MPI_Wait(&req[index], &status);
/* check success code from process */
if (flags[index + w] != SCR_SUCCESS) {
success = SCR_FAILURE;
}
/* one less request outstanding now */
outstanding--;
}
/* free the MPI_Request arrays */
scr_free(&req);
scr_free(&flags);
} else {
/* wait for start signal from rank 0 */
MPI_Status status;
MPI_Recv(&success, 1, MPI_INT, 0, 0, scr_comm_world, &status);
/* if rank 0 hasn't seen a failure, try to read in our files */
if (success == SCR_SUCCESS) {
/* fetch these files into the directory */
if (scr_fetch_files_list(file_list, dir, map) != SCR_SUCCESS) {
success = SCR_FAILURE;
}
}
/* tell rank 0 that we're done and send him our success code */
MPI_Send(&success, 1, MPI_INT, 0, 0, scr_comm_world);
}
/* determine whether all processes successfully read their files */
if (scr_alltrue(success == SCR_SUCCESS)) {
return SCR_SUCCESS;
}
return SCR_FAILURE;
}
/* apply XOR redundancy scheme to dataset files */
static int scr_reddesc_apply_xor(scr_filemap* map, const scr_reddesc* c, int id)
{
int rc = SCR_SUCCESS;
int i;
/* get pointer to XOR state structure */
scr_reddesc_xor* state = (scr_reddesc_xor*) c->copy_state;
/* allocate buffer to read a piece of my file */
char* send_buf = (char*) scr_align_malloc(scr_mpi_buf_size, scr_page_size);
if (send_buf == NULL) {
scr_abort(-1, "Allocating memory for send buffer: malloc(%d) errno=%d %s @ %s:%d",
scr_mpi_buf_size, errno, strerror(errno), __FILE__, __LINE__
);
}
/* allocate buffer to read a piece of the recevied chunk file */
char* recv_buf = (char*) scr_align_malloc(scr_mpi_buf_size, scr_page_size);
if (recv_buf == NULL) {
scr_abort(-1, "Allocating memory for recv buffer: malloc(%d) errno=%d %s @ %s:%d",
scr_mpi_buf_size, errno, strerror(errno), __FILE__, __LINE__
);
}
/* count the number of files I have and allocate space in structures for each of them */
int num_files = scr_filemap_num_files(map, id, scr_my_rank_world);
int* fds = (int*) SCR_MALLOC(num_files * sizeof(int));
char** filenames = (char**) SCR_MALLOC(num_files * sizeof(char*));
unsigned long* filesizes = (unsigned long*) SCR_MALLOC(num_files * sizeof(unsigned long));
/* record partner's redundancy descriptor hash in our filemap */
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);
/* allocate a new xor file header hash */
scr_hash* header = scr_hash_new();
/* record the global ranks of the processes in our xor group */
scr_hash_merge(header, state->group_map);
/* record dataset in header */
scr_hash* dataset = scr_hash_new();
scr_filemap_get_dataset(map, id, scr_my_rank_world, dataset);
scr_hash_set(header, SCR_KEY_COPY_XOR_DATASET, dataset);
/* open each file, get the filesize of each, and read the meta data of each */
scr_hash* current_files = scr_hash_new();
int file_count = 0;
unsigned long my_bytes = 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 */
filenames[file_count] = scr_hash_elem_key(file_elem);
/* get the filesize of this file and add the byte count to the total */
filesizes[file_count] = scr_file_size(filenames[file_count]);
my_bytes += filesizes[file_count];
/* read the meta data for this file and insert it into the current_files hash */
scr_meta* file_hash = scr_meta_new();
scr_filemap_get_meta(map, id, scr_my_rank_world, filenames[file_count], file_hash);
scr_hash_setf(current_files, file_hash, "%d", file_count);
/* open the file */
fds[file_count] = scr_open(filenames[file_count], O_RDONLY);
if (fds[file_count] < 0) {
/* TODO: try again? */
scr_abort(-1, "Opening checkpoint file for copying: scr_open(%s, O_RDONLY) errno=%d %s @ %s:%d",
filenames[file_count], errno, strerror(errno), __FILE__, __LINE__
);
}
file_count++;
}
/* set total number of files we have, plus our rank */
scr_hash* current_hash = scr_hash_new();
scr_hash_set_kv_int(current_hash, SCR_KEY_COPY_XOR_RANK, scr_my_rank_world);
scr_hash_set_kv_int(current_hash, SCR_KEY_COPY_XOR_FILES, file_count);
scr_hash_set(current_hash, SCR_KEY_COPY_XOR_FILE, current_files);
/* exchange file info with partners and add data to our header */
scr_hash* partner_hash = scr_hash_new();
scr_hash_sendrecv(current_hash, state->rhs_rank, partner_hash, state->lhs_rank, c->comm);
scr_hash_set(header, SCR_KEY_COPY_XOR_CURRENT, current_hash);
scr_hash_set(header, SCR_KEY_COPY_XOR_PARTNER, partner_hash);
/* allreduce to get maximum filesize */
unsigned long max_bytes;
MPI_Allreduce(&my_bytes, &max_bytes, 1, MPI_UNSIGNED_LONG, MPI_MAX, c->comm);
/* TODO: use unsigned long integer arithmetic (with proper byte padding) instead of char to speed things up */
/* compute chunk size according to maximum file length and number of ranks in xor set */
/* if filesize doesn't divide evenly, then add one byte to chunk_size */
/* TODO: check that ranks > 1 for this divide to be safe (or at partner selection time) */
size_t chunk_size = max_bytes / (unsigned long) (c->ranks - 1);
if ((c->ranks - 1) * chunk_size < max_bytes) {
chunk_size++;
}
/* TODO: need something like this to handle 0-byte files? */
if (chunk_size == 0) {
chunk_size++;
}
/* record the dataset id and the chunk size in the xor chunk header */
scr_hash_util_set_bytecount(header, SCR_KEY_COPY_XOR_CHUNK, chunk_size);
/* set chunk filenames of form: xor.<group_id>_<xor_rank+1>_of_<xor_ranks>.scr */
char my_chunk_file[SCR_MAX_FILENAME];
char* dir = scr_cache_dir_hidden_get(c, id);
sprintf(my_chunk_file, "%s/xor.%d_%d_of_%d.scr", dir, c->group_id, c->rank+1, c->ranks);
scr_free(&dir);
/* record chunk file in filemap before creating it */
scr_filemap_add_file(map, id, scr_my_rank_world, my_chunk_file);
scr_filemap_write(scr_map_file, map);
/* open my chunk file */
mode_t mode_file = scr_getmode(1, 1, 0);
int fd_chunk = scr_open(my_chunk_file, O_WRONLY | O_CREAT | O_TRUNC, mode_file);
if (fd_chunk < 0) {
/* TODO: try again? */
scr_abort(-1, "Opening XOR chunk file for writing: scr_open(%s) errno=%d %s @ %s:%d",
my_chunk_file, errno, strerror(errno), __FILE__, __LINE__
);
}
/* write out the xor chunk header */
scr_hash_write_fd(my_chunk_file, fd_chunk, header);
scr_hash_delete(&header);
MPI_Request request[2];
MPI_Status status[2];
/* XOR Reduce_scatter */
size_t nread = 0;
while (nread < chunk_size) {
size_t count = chunk_size - nread;
if (count > scr_mpi_buf_size) {
count = scr_mpi_buf_size;
}
int chunk_id;
for(chunk_id = c->ranks-1; chunk_id >= 0; chunk_id--) {
/* read the next set of bytes for this chunk from my file into send_buf */
if (chunk_id > 0) {
int chunk_id_rel = (c->rank + c->ranks + chunk_id) % c->ranks;
if (chunk_id_rel > c->rank) {
chunk_id_rel--;
}
unsigned long offset = chunk_size * (unsigned long) chunk_id_rel + nread;
if (scr_read_pad_n(num_files, filenames, fds,
send_buf, count, offset, filesizes) != SCR_SUCCESS)
{
rc = SCR_FAILURE;
}
} else {
memset(send_buf, 0, count);
}
/* TODO: XORing with unsigned long would be faster here (if chunk size is multiple of this size) */
/* merge the blocks via xor operation */
if (chunk_id < c->ranks-1) {
for (i = 0; i < count; i++) {
send_buf[i] ^= recv_buf[i];
}
}
if (chunk_id > 0) {
/* not our chunk to write, forward it on and get the next */
MPI_Irecv(recv_buf, count, MPI_BYTE, state->lhs_rank, 0, c->comm, &request[0]);
MPI_Isend(send_buf, count, MPI_BYTE, state->rhs_rank, 0, c->comm, &request[1]);
MPI_Waitall(2, request, status);
} else {
/* write send block to send chunk file */
if (scr_write_attempt(my_chunk_file, fd_chunk, send_buf, count) != count) {
rc = SCR_FAILURE;
}
}
}
nread += count;
}
/* close my chunkfile, with fsync */
if (scr_close(my_chunk_file, fd_chunk) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* close my dataset files */
for (i=0; i < num_files; i++) {
scr_close(filenames[i], fds[i]);
}
/* free the buffers */
scr_free(&filesizes);
/* in this case, we don't free each name, since we copied the pointer to the string in the filemap */
scr_free(&filenames);
scr_free(&fds);
scr_align_free(&send_buf);
scr_align_free(&recv_buf);
/* TODO: need to check for errors */
/* write meta file for xor chunk */
unsigned long my_chunk_file_size = scr_file_size(my_chunk_file);
scr_meta* meta = scr_meta_new();
scr_meta_set_filename(meta, my_chunk_file);
scr_meta_set_filetype(meta, SCR_META_FILE_XOR);
scr_meta_set_filesize(meta, my_chunk_file_size);
scr_meta_set_complete(meta, 1);
/* TODODSET: move the ranks field elsewhere, for now it's needed by scr_index.c */
scr_meta_set_ranks(meta, scr_ranks_world);
scr_filemap_set_meta(map, id, scr_my_rank_world, my_chunk_file, meta);
scr_filemap_write(scr_map_file, map);
scr_meta_delete(&meta);
/* if crc_on_copy is set, compute and store CRC32 value for chunk file */
if (scr_crc_on_copy) {
scr_compute_crc(map, id, scr_my_rank_world, my_chunk_file);
/* TODO: would be nice to save this CRC in our partner's XOR file so we can check correctness on a rebuild */
}
return rc;
}
