/*
* Sets up local data structure for series of reads on an adios file
* - create evpath graph and structures
* -- create evpath control stone (outgoing)
* -- create evpath data stone (incoming)
* -- rank 0 dumps contact info to file
* -- create connections using contact info from file
*/
ADIOS_FILE*
adios_read_flexpath_open(const char * fname,
MPI_Comm comm,
enum ADIOS_LOCKMODE lock_mode,
float timeout_sec)
{
fp_log("FUNC", "entering flexpath_open\n");
ADIOS_FILE *adiosfile = malloc(sizeof(ADIOS_FILE));
if(!adiosfile){
adios_error (err_no_memory,
"Cannot allocate memory for file info.\n");
return NULL;
}
flexpath_reader_file *fp = new_flexpath_reader_file(fname);
adios_errno = 0;
fp->stone = EValloc_stone(fp_read_data->fp_cm);
fp->comm = comm;
MPI_Comm_size(fp->comm, &(fp->size));
MPI_Comm_rank(fp->comm, &(fp->rank));
EVassoc_terminal_action(fp_read_data->fp_cm,
fp->stone,
op_format_list,
op_msg_handler,
adiosfile);
EVassoc_terminal_action(fp_read_data->fp_cm,
fp->stone,
update_step_msg_format_list,
update_step_msg_handler,
adiosfile);
EVassoc_terminal_action(fp_read_data->fp_cm,
fp->stone,
evgroup_format_list,
group_msg_handler,
adiosfile);
EVassoc_raw_terminal_action(fp_read_data->fp_cm,
fp->stone,
raw_handler,
adiosfile);
/* Gather the contact info from the other readers
and write it to a file. Create a ready file so
that the writer knows it can parse this file. */
double setup_start = dgettimeofday();
char writer_ready_filename[200];
char writer_info_filename[200];
char reader_ready_filename[200];
char reader_info_filename[200];
sprintf(reader_ready_filename, "%s_%s", fname, READER_READY_FILE);
sprintf(reader_info_filename, "%s_%s", fname, READER_CONTACT_FILE);
sprintf(writer_ready_filename, "%s_%s", fname, WRITER_READY_FILE);
sprintf(writer_info_filename, "%s_%s", fname, WRITER_CONTACT_FILE);
char *string_list;
char data_contact_info[CONTACT_LENGTH];
string_list = attr_list_to_string(CMget_contact_list(fp_read_data->fp_cm));
sprintf(&data_contact_info[0], "%d:%s", fp->stone, string_list);
free(string_list);
char * recvbuf;
if(fp->rank == 0){
recvbuf = (char*)malloc(sizeof(char)*CONTACT_LENGTH*(fp->size));
}
MPI_Gather(data_contact_info, CONTACT_LENGTH, MPI_CHAR, recvbuf,
CONTACT_LENGTH, MPI_CHAR, 0, fp->comm);
if(fp->rank == 0){
// print our own contact information
FILE * fp_out = fopen(reader_info_filename, "w");
int i;
if(!fp_out){
adios_error(err_file_open_error,
"File for contact info could not be opened for writing.\n");
exit(1);
}
for(i=0; i<fp->size; i++) {
fprintf(fp_out,"%s\n", &recvbuf[i*CONTACT_LENGTH]);
}
fclose(fp_out);
free(recvbuf);
FILE * read_ready = fopen(reader_ready_filename, "w");
fprintf(read_ready, "ready");
fclose(read_ready);
}
MPI_Barrier(fp->comm);
FILE * fp_in = fopen(writer_ready_filename,"r");
while(!fp_in) {
//CMsleep(fp_read_data->fp_cm, 1);
fp_in = fopen(writer_ready_filename, "r");
}
fclose(fp_in);
fp_in = fopen(writer_info_filename, "r");
while(!fp_in){
//CMsleep(fp_read_data->fp_cm, 1);
fp_in = fopen(writer_info_filename, "r");
}
char in_contact[CONTACT_LENGTH] = "";
//fp->bridges = malloc(sizeof(bridge_info));
int num_bridges = 0;
int their_stone;
// change to read all numbers, dont create stones, turn bridge array into linked list
while(fscanf(fp_in, "%d:%s", &their_stone, in_contact) != EOF){
//fprintf(stderr, "writer contact: %d:%s\n", their_stone, in_contact);
fp->bridges = realloc(fp->bridges,
sizeof(bridge_info) * (num_bridges+1));
fp->bridges[num_bridges].their_num = their_stone;
fp->bridges[num_bridges].contact = strdup(in_contact);
fp->bridges[num_bridges].created = 0;
fp->bridges[num_bridges].step = 0;
fp->bridges[num_bridges].opened = 0;
fp->bridges[num_bridges].scheduled = 0;
num_bridges++;
}
fclose(fp_in);
fp->num_bridges = num_bridges;
// clean up of writer's files
MPI_Barrier(fp->comm);
if(fp->rank == 0){
unlink(writer_info_filename);
unlink(writer_ready_filename);
}
adiosfile->fh = (uint64_t)fp;
adiosfile->current_step = 0;
/* Init with a writer to get initial scalar
data so we can handle inq_var calls and
also populate the ADIOS_FILE struct. */
double bridge_start = MPI_Wtime();
if(fp->size < num_bridges){
int mystart = (num_bridges/fp->size) * fp->rank;
int myend = (num_bridges/fp->size) * (fp->rank+1);
fp->writer_coordinator = mystart;
int z;
for(z=mystart; z<myend; z++){
build_bridge(&fp->bridges[z]);
}
}
else{
int writer_rank = fp->rank % num_bridges;
build_bridge(&fp->bridges[writer_rank]);
fp->writer_coordinator = writer_rank;
}
// requesting initial data.
send_open_msg(fp, fp->writer_coordinator);
fp->data_read = 0;
send_flush_msg(fp, fp->writer_coordinator, DATA, 1);
send_flush_msg(fp, fp->writer_coordinator, EVGROUP, 1);
fp->data_read = 0;
// this has to change. Writer needs to have some way of
// taking the attributes out of the xml document
// and sending them over ffs encoded. Not yet implemented.
// the rest of this info for adiosfile gets filled in raw_handler.
adiosfile->nattrs = 0;
adiosfile->attr_namelist = NULL;
// first step is at least one, otherwise raw_handler will not execute.
// in reality, writer might be further along, so we might have to make
// the writer explitly send across messages each time it calls close, to
// indicate which timesteps are available.
adiosfile->last_step = 1;
adiosfile->path = strdup(fname);
// verifies these two fields. It's not BP, so no BP version.
// It's a stream, so how can the file size be known?
adiosfile->version = -1;
adiosfile->file_size = 0;
adios_errno = err_no_error;
fp_log("FUNC", "leaving flexpath_open\n");
return adiosfile;
}
int
adios_read_icee_init_method (MPI_Comm comm, PairStruct* params)
{
log_debug ("%s\n", __FUNCTION__);
int cm_port = 59997;
char *cm_host = "localhost";
int cm_remote_port = 59999;
char *cm_remote_host = "localhost";
char *cm_attr = NULL;
//attr_list contact_list;
icee_transport_t icee_transport_init = TCP;
icee_transport_t icee_transport = TCP;
icee_contactinfo_rec_t *remote_contact = NULL;
int i;
int use_single_remote_server = 1;
char *remote_list_str = NULL;
char *attr_list_str = NULL;
int use_native_contact = 0;
PairStruct * p = params;
while (p)
{
if (!strcasecmp (p->name, "cm_attr"))
{
cm_attr = p->value;
}
else if (!strcasecmp (p->name, "cm_host"))
{
cm_host = p->value;
}
else if (!strcasecmp (p->name, "cm_port"))
{
cm_port = atoi(p->value);
}
else if (!strcasecmp (p->name, "cm_remote_host"))
{
cm_remote_host = p->value;
}
else if (!strcasecmp (p->name, "cm_remote_port"))
{
cm_remote_port = atoi(p->value);
}
else if (!strcasecmp (p->name, "remote_list"))
{
use_single_remote_server = 0;
if (p->value)
remote_list_str = strdup(p->value);
}
else if (!strcasecmp (p->name, "attr_list"))
{
use_single_remote_server = 0;
if (p->value)
attr_list_str = strdup(p->value);
}
else if (!strcasecmp (p->name, "transport"))
{
if (strcasecmp(p->value, "TCP") == 0)
icee_transport = TCP;
else if (strcasecmp(p->value, "ENET") == 0)
icee_transport = ENET;
else if (strcasecmp(p->value, "NNTI") == 0)
icee_transport = NNTI;
else if (strcasecmp(p->value, "IB") == 0)
icee_transport = IB;
else
log_error ("No support: %s\n", p->value);
}
else if (!strcasecmp (p->name, "transport_init"))
{
if (strcasecmp(p->value, "TCP") == 0)
icee_transport_init = TCP;
else if (strcasecmp(p->value, "ENET") == 0)
icee_transport_init = ENET;
else if (strcasecmp(p->value, "NNTI") == 0)
icee_transport_init = NNTI;
else if (strcasecmp(p->value, "IB") == 0)
icee_transport_init = IB;
else
log_error ("No support: %s\n", p->value);
}
else if (!strcasecmp (p->name, "num_parallel"))
{
icee_read_num_parallel = atoi(p->value);
}
else if (!strcasecmp (p->name, "is_passive"))
{
is_read_cm_passive = atoi(p->value);
}
else if (!strcasecmp (p->name, "use_native_contact"))
{
use_native_contact = atoi(p->value);
}
p = p->next;
}
pthread_mutex_init(&fileinfo_lock, NULL);
if (use_single_remote_server)
{
num_remote_server = 1;
attr_list contact_list = create_attr_list();
set_contact_list(contact_list, icee_transport_init, cm_remote_host, cm_remote_port);
icee_contactinfo_rec_t *p;
p = malloc(sizeof(icee_contactinfo_rec_t));
char *contact_string = attr_list_to_string(contact_list);
p->contact_string = contact_string;
p->stone_id = 0; // we assume. it can be wrong.
p->next = NULL;
remote_contact = p;
}
else
{
num_remote_server = 0;
icee_contactinfo_rec_t *p;
icee_contactinfo_rec_t *prev;
char* token = strtok(remote_list_str, ",");
while (token)
{
char host[256];
int port = 0;
if (token[0] == ':')
{
strcpy(host, cm_remote_host);
port = atoi(token+1);
}
else
{
char *pch = strchr(token, ':');
if (pch != NULL)
{
strncpy(host, token, pch - token);
host[pch-token] = '\0';
port = atoi(pch+1);
}
else
{
int len = strlen(token);
strncpy(host, token, len);
assert(len < 256);
host[len] = '\0';
port = cm_remote_port;
}
}
log_debug("Remote server list: (%d) %s:%d\n", num_remote_server, host, port);
p = malloc(sizeof(icee_contactinfo_rec_t));
attr_list contact_list;
contact_list = create_attr_list();
set_contact_list(contact_list, icee_transport_init, host, port);
p->contact_string = attr_list_to_string(contact_list);
p->stone_id = 0; // we assume. it can be wrong.
p->next = NULL;
if (num_remote_server == 0)
remote_contact = p;
else
prev->next = p;
prev = p;
num_remote_server++;
token = strtok(NULL, ",");
}
}
if (attr_list_str != NULL)
{
num_remote_server = 0;
icee_contactinfo_rec_t *p;
icee_contactinfo_rec_t *prev;
char* token = strtok(attr_list_str, ",");
while (token)
{
int remote_stone = 0;
char string_list[256];
sscanf(token, "%d:%s", &remote_stone, &string_list[0]);
p = malloc(sizeof(icee_contactinfo_rec_t));
attr_list contact_list;
p->stone_id = remote_stone;
p->contact_string = strdup(string_list);
p->next = NULL;
if (num_remote_server == 0)
remote_contact = p;
else
prev->next = p;
prev = p;
num_remote_server++;
token = strtok(NULL, ",");
}
}
if (icee_read_num_parallel > ICEE_MAX_PARALLEL)
{
icee_read_num_parallel = ICEE_MAX_PARALLEL;
log_info ("Max. number of threads is set to %d\n", icee_read_num_parallel);
}
log_debug ("transport : %s\n", icee_transport_name[icee_transport]);
/*
log_info ("cm_host : %s\n", cm_host);
log_info ("cm_port : %d\n", cm_port);
for (i = 0; i < num_remote_server; i++)
{
log_info ("remote_list : %s:%d\n", remote_server[i].client_host, remote_server[i].client_port);
}
*/
if (!adios_read_icee_initialized)
{
if (is_read_cm_passive)
{
icee_contactinfo_rec_t *prev;
for (i = 0; i < num_remote_server; i++)
{
attr_list contact_list;
icee_contactinfo_rec_t *p = (i == 0)? remote_contact : prev->next;
pcm[i] = CManager_create();
if (!CMfork_comm_thread(pcm[i]))
printf("Fork of communication thread[%d] failed.\n", i);
contact_list = attr_list_from_string(p->contact_string);
log_debug("Passive remote contact: \"%s\"\n", attr_list_to_string(contact_list));
if (adios_verbose_level > 5) dump_attr_list(contact_list);
/*
attr_list contact_list = create_attr_list();
add_string_attr(contact_list, attr_atom_from_string("IP_HOST"),
remote_server[i].client_host);
add_int_attr(contact_list, attr_atom_from_string("IP_PORT"),
remote_server[i].client_port);
*/
CMConnection conn = CMinitiate_conn(pcm[i], contact_list);
int n = 0;
while (conn == NULL)
{
log_error ("Passive connection failed (%d). Try again ...\n", i);
dump_attr_list(contact_list);
sleep(2);
conn = CMinitiate_conn(pcm[i], contact_list);
if (n > 5) break;
n++;
}
if (conn == NULL)
{
log_error ("Initializing passive connection failed (%d)\n", i);
}
CMFormat fm_checkin, fm_fileinfo;
fm_checkin = CMregister_format(pcm[i], icee_passivecheckin_format_list);
CMregister_handler(fm_checkin, icee_passivecheckin_reply_handler, on_icee_passivecheckin_reply);
fm_fileinfo = CMregister_format(pcm[i], icee_fileinfo_format_list);
CMregister_handler(fm_fileinfo, icee_fileinfo_recv_handler, on_icee_fileinfo_recv);
icee_passivecheckin_rec_t m;
int condition = CMCondition_get(pcm[i], conn);
CMCondition_set_client_data(pcm[i], condition, NULL);
m.condition = condition;
if (CMwrite(conn, fm_checkin, (void*)&m) != 1)
log_error ("Passive check-in failed (%d)\n", i);
prev = p;
}
log_debug("Passive connection established");
goto done;
}
EVstone stone[ICEE_MAX_PARALLEL], remote_stone;
EVsource source;
attr_list contact[ICEE_MAX_PARALLEL];
icee_contactinfo_rec_t contact_msg[ICEE_MAX_PARALLEL];
for (i=0; i<icee_read_num_parallel; i++)
{
icee_read_cm[i] = CManager_create();
contact[i] = create_attr_list();
set_contact_list(contact[i], icee_transport, cm_host, cm_port+i);
if (CMlisten_specific(icee_read_cm[i], contact[i]) == 0)
printf("Error: unable to initialize connection manager[%d].\n", i);
if (!CMfork_comm_thread(icee_read_cm[i]))
printf("Fork of communication thread[%d] failed.\n", i);
stone[i] = EValloc_stone(icee_read_cm[i]);
if (adios_verbose_level > 5)
{
log_debug("Reader contact: \"%d:%s\"\n", stone[i], attr_list_to_string(CMget_contact_list(icee_read_cm[i])));
dump_attr_list(CMget_contact_list(icee_read_cm[i]));
}
EVassoc_terminal_action(icee_read_cm[i], stone[i], icee_fileinfo_format_list, icee_fileinfo_handler, NULL);
contact_msg[i].stone_id = stone[i];
attr_list contact_list;
if (use_native_contact)
contact_list = CMget_contact_list(icee_read_cm[i]);
else
contact_list = contact[i];
contact_msg[i].contact_string = attr_list_to_string(contact_list);
contact_msg[i].next = NULL;
if (i>0)
contact_msg[i-1].next = &contact_msg[i];
}
EVstone split_stone;
EVaction split_action;
split_stone = EValloc_stone(icee_read_cm[0]);
split_action = EVassoc_split_action(icee_read_cm[0], split_stone, NULL);
icee_contactinfo_rec_t *prev;
for (i = 0; i < num_remote_server; i++)
{
attr_list contact_list;
EVstone remote_stone, output_stone;
output_stone = EValloc_stone(icee_read_cm[0]);
icee_contactinfo_rec_t *p = (i == 0)? remote_contact : prev->next;
remote_stone = p->stone_id;
contact_list = attr_list_from_string(p->contact_string);
EVaction action;
action = EVassoc_bridge_action(icee_read_cm[0], output_stone, contact_list, remote_stone);
int n = 0;
while (action == -1)
{
log_error ("Connection failed (%d). Try again ...\n", i);
dump_attr_list(contact_list);
sleep(2);
action = EVassoc_bridge_action(icee_read_cm[0], output_stone, contact_list, remote_stone);
if (n > 5) break;
n++;
}
EVaction_add_split_target(icee_read_cm[0], split_stone, split_action, output_stone);
prev = p;
log_debug("Remote contact: \"%d:%s\"\n", remote_stone, attr_list_to_string(contact_list));
if (adios_verbose_level > 5) dump_attr_list(contact_list);
}
source = EVcreate_submit_handle(icee_read_cm[0], split_stone, icee_contactinfo_format_list);
//if (adios_verbose_level > 5) icee_contactinfo_print(contact_msg);
EVsubmit(source, contact_msg, NULL);
done:
adios_read_icee_initialized = 1;
}
return 0;
}