int do_KLR_mapping(network_t *network, options_t *opt, idx_t *part) {
int ret = ORCC_OK;
int i;
proc_info_t *processors;
assert(network != NULL);
assert(opt != NULL);
assert(part != NULL);
processors = init_processors(opt->nb_processors);
print_orcc_trace(ORCC_VL_VERBOSE_1, "Applying Kernighan Lin Refinement Weighted strategy for mapping");
sort_actors(network->actors, network->nb_actors);
for (i = 0; i < network->nb_actors; i++) {
assign_actor_to_min_utilized_processor(network, part, processors, opt->nb_processors, i);
}
do_KL_algorithm(network, part, processors, opt->nb_processors);
if (check_verbosity(ORCC_VL_VERBOSE_2) == TRUE) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : KLRLB result");
for (i = 0; i < network->nb_actors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Actor[%d]\tname = %s\tworkload = %d\tprocessorId = %d",
i, network->actors[i]->name, network->actors[i]->workload, network->actors[i]->processor_id);
}
}
delete_processors(processors);
return ret;
}
/**
* Round Robin strategy
* @author Long Nguyen
*/
int do_round_robbin_mapping(network_t *network, options_t *opt, idx_t *part) {
int ret = ORCC_OK;
int i, k = 0;
assert(network != NULL);
assert(opt != NULL);
assert(part != NULL);
print_orcc_trace(ORCC_VL_VERBOSE_1, "Applying Round Robin strategy for mapping");
sort_actors(network->actors, network->nb_actors);
for (i = 0; i < network->nb_actors; i++) {
network->actors[i]->processor_id = k++;
part[i] = network->actors[i]->processor_id;
// There must be something needing to be improved here, i.e. invert
// the direction of the distribution to have more balancing.
if (k >= opt->nb_processors)
k = 0;
}
if (check_verbosity(ORCC_VL_VERBOSE_2) == TRUE) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Round Robin result");
for (i = 0; i < network->nb_actors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Actor[%d]\tname = %s\tworkload = %d\tprocessorId = %d",
i, network->actors[i]->name, network->actors[i]->workload, network->actors[i]->processor_id);
}
}
return ret;
}
int do_weighted_round_robin_comm_mapping(network_t *network, options_t *opt, idx_t *part) {
int ret = ORCC_OK;
int i, j;
int selectedProc, minCommIndex;
proc_info_t *processors;
assert(network != NULL);
assert(opt != NULL);
assert(part != NULL);
print_orcc_trace(ORCC_VL_VERBOSE_1, "Applying Communication Optimized Weighted strategy for mapping");
processors = init_processors(opt->nb_processors);
sort_actors(network->actors, network->nb_actors);
for (i = 0; i < network->nb_actors; i++) {
selectedProc = find_min_utilized_processor(processors, opt->nb_processors);
minCommIndex = -1;
for (j = 0; j < network->nb_actors; j++) {
if (network->actors[j]->processor_id == -1) {
if (minCommIndex == -1) {
minCommIndex = j;
}
if (calculate_comm_of_actor(network, processors, j, selectedProc)
<= calculate_comm_of_actor(network, processors, minCommIndex, selectedProc)) {
minCommIndex = j;
}
}
}
network->actors[minCommIndex]->processor_id = processors[selectedProc].processor_id;
part[minCommIndex] = network->actors[minCommIndex]->processor_id;
processors[selectedProc].utilization += network->actors[minCommIndex]->workload;
}
if (check_verbosity(ORCC_VL_VERBOSE_2) == TRUE) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : COWLB result");
for (i = 0; i < network->nb_actors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Actor[%d]\tname = %s\tworkload = %d\tprocessorId = %d",
i, network->actors[i]->name, network->actors[i]->workload, network->actors[i]->processor_id);
}
for (i = 0; i < opt->nb_processors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Workload Proc[%d]: %d",
i, processors[i].utilization);
}
}
delete_processors(processors);
return ret;
}
int sort_actors(actor_t **actors, int nb_actors) {
int ret = ORCC_OK;
int i, j;
assert(actors != NULL);
for (i = 0; i < nb_actors; i++) {
for (j = 0; j < nb_actors - i - 1; j++) {
if (actors[j]->workload <= actors[j+1]->workload) {
swap_actors(actors, j, j+1, nb_actors);
}
}
}
if (check_verbosity(ORCC_VL_VERBOSE_2) == TRUE) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : The sorted list:");
for (i = 0; i < nb_actors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Actor[%d]\tid = %s\tworkload = %d", i, actors[i]->name, actors[i]->workload);
}
}
return ret;
}
/**
* Entry point for all mapping strategies
*/
int do_mapping(network_t *network, options_t *opt, mapping_t *mapping) {
int ret = ORCC_OK;
idx_t *part;
ticks startTime, endTime;
assert(network != NULL);
assert(opt != NULL);
assert(mapping != NULL);
part = (idx_t*) malloc(sizeof(idx_t) * (network->nb_actors));
if(check_verbosity(ORCC_VL_VERBOSE_2)) {
print_network(network);
}
startTime = getticks();
if (opt->nb_processors != 1) {
switch (opt->strategy) {
#ifdef METIS_ENABLE
case ORCC_MS_METIS_REC:
ret = do_metis_recursive_partition(network, opt, part);
break;
case ORCC_MS_METIS_KWAY_CV:
ret = do_metis_kway_partition(network, opt, part, METIS_OBJTYPE_CUT); /*TODO : should be METIS_OBJTYPE_VOL : Metis seem's to invert its options */
break;
case ORCC_MS_METIS_KWAY_EC:
ret = do_metis_kway_partition(network, opt, part, METIS_OBJTYPE_VOL); /*TODO : should be METIS_OBJTYPE_CUT : Metis seem's to invert its options */
break;
#endif
case ORCC_MS_ROUND_ROBIN:
ret = do_round_robbin_mapping(network, opt, part);
break;
case ORCC_MS_QM:
ret = do_quick_mapping(network, opt, part);
break;
case ORCC_MS_WLB:
ret = do_weighted_round_robin_mapping(network, opt, part);
break;
case ORCC_MS_COWLB:
ret = do_weighted_round_robin_comm_mapping(network, opt, part);
break;
case ORCC_MS_KRWLB:
ret = do_KLR_mapping(network, opt, part);
break;
default:
break;
}
} else {
int i;
for (i = 0; i < network->nb_actors; i++) {
part[i] = 0;
}
}
endTime = getticks();
set_mapping_from_partition(network, part, mapping);
if(check_verbosity(ORCC_VL_VERBOSE_1)) {
print_mapping(mapping);
print_load_balancing(mapping);
print_edge_cut(network);
print_orcc_trace(ORCC_VL_VERBOSE_2, "Mapping time : %2.lf", elapsed(endTime, startTime));
}
free(part);
return ret;
}
/* Parse a single option. */
static error_t parse_opt(int key, char *opt_arg, struct argp_state *state)
{
/* Get the input argument from argp_parse, which we
know is a pointer to our arguments structure. */
arguments *arg = state->input;
switch (key) {
case 'q':
check_verbosity(state, arg, orcm_quiet);
arg->verbosity = orcm_quiet;
break;
case 'v':
check_verbosity(state, arg, orcm_verbose);
arg->verbosity = orcm_verbose;
break;
case 'd':
check_verbosity(state, arg, orcm_debug);
arg->verbosity = orcm_debug;
break;
case 'c':
check_color(state,arg, orcc_use_color);
arg->color = orcc_use_color;
break;
case 'n':
check_color(state,arg, orcc_no_color);
arg->color = orcc_no_color;
break;
case 'e':
if(1 != sscanf(opt_arg, "%d", &(arg->max_events))) {
orcerror("Events set to a non integer value.\n");
argp_usage(state);
}
if (1 > arg->max_events) {
orcerror("Events set to a 0 or a negative value.\n");
argp_usage(state);
}
break;
case 1001:
arg->excludes_len++;
size_t size = arg->excludes_len * sizeof(*(arg->excludes));
char **tmp;
tmp = realloc(arg->excludes, size);
if (0 == tmp) {
if (0 != arg->excludes) {
free(arg->excludes);
}
orcerror("%s (%d)\n", strerror(errno), errno);
exit(EXIT_FAILURE);
}
tmp[arg->excludes_len - 1] = opt_arg;
arg->excludes = tmp;
break;
case 'o':
arg->out_file = opt_arg;
break;
case ARGP_KEY_ARG:
if (state->arg_num > 1) {
/* Too many arguments. */
argp_usage(state);
}
if (('h'== opt_arg[0] && 't' == opt_arg[1] &&'t' == opt_arg[2] &&
'p' == opt_arg[3] && ':' == opt_arg[4] && '/' == opt_arg[5] &&
'/' == opt_arg[6]) ||
('h'== opt_arg[0] && 't' == opt_arg[1] &&'t' == opt_arg[2] &&
'p' == opt_arg[3] && 's' == opt_arg[4] && ':' == opt_arg[5] &&
'/' == opt_arg[6] && '/'== opt_arg[7]))
{
arg->url = opt_arg;
} else {
orcerror("Add http:// or https:// to your target url.\n",
strerror(errno), errno);
exit(EXIT_FAILURE);
}
break;
case ARGP_KEY_END:
if (state->arg_num < 1) {
/* Not enough arguments. */
argp_usage(state);
}
break;
default:
return ARGP_ERR_UNKNOWN;
}
return 0;
}
int do_quick_mapping(network_t *network, options_t *opt, idx_t *part) {
int ret = ORCC_OK;
int i, unMappedActors, selectedProcIndex, maxCommCost, maxIndex, total_workload = 0;
proc_info_t *processors;
assert(network != NULL);
assert(opt != NULL);
assert(part != NULL);
processors = init_processors(opt->nb_processors);
print_orcc_trace(ORCC_VL_VERBOSE_1, "Applying Quick Mapping strategy for mapping");
unMappedActors = network->nb_actors;
selectedProcIndex = 0;
for (i = 0; i < network->nb_actors; i++) {
total_workload += network->actors[i]->workload;
}
while (unMappedActors > 0) {
maxIndex = 0;
maxCommCost = 0;
for (i = 0; i < network->nb_actors; i++) {
if (network->actors[i]->processor_id == -1) {
if (network->actors[i]->triedProcId != processors[selectedProcIndex].processor_id) {
network->actors[i]->commCost = 0;
network->actors[i]->triedProcId = processors[selectedProcIndex].processor_id;
}
if (network->actors[i]->commCost >= maxCommCost) {
maxCommCost = network->actors[i]->commCost;
maxIndex = i;
}
}
}
network->actors[maxIndex]->processor_id = processors[selectedProcIndex].processor_id;
part[maxIndex] = network->actors[maxIndex]->processor_id;
processors[selectedProcIndex].utilization += network->actors[maxIndex]->workload;
for (i = 0; i < network->nb_connections; i++) {
if (network->connections[i]->src->id == network->actors[maxIndex]->id) {
network->actors[network->connections[i]->dst->id]->commCost += network->connections[i]->workload;
}
if (network->connections[i]->dst->id == network->actors[maxIndex]->id) {
network->actors[network->connections[i]->src->id]->commCost += network->connections[i]->workload;
}
}
if (processors[selectedProcIndex].utilization >= total_workload / opt->nb_processors
&& selectedProcIndex < opt->nb_processors - 1) {
selectedProcIndex++;
}
unMappedActors--;
}
if (check_verbosity(ORCC_VL_VERBOSE_2) == TRUE) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : QM result");
for (i = 0; i < network->nb_actors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Actor[%d]\tname = %s\tworkload = %d\tprocessorId = %d",
i, network->actors[i]->name, network->actors[i]->workload, network->actors[i]->processor_id);
}
for (i = 0; i < opt->nb_processors; i++) {
print_orcc_trace(ORCC_VL_VERBOSE_2, "DEBUG : Workload Proc[%d]: %d",
i, processors[i].utilization);
}
}
delete_processors(processors);
return ret;
}
