/* collective operation for the torus network */
void torus_collective(char* category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
{
tw_event * e_new;
tw_stime xfer_to_nic_time;
nodes_message * msg;
tw_lpid local_nic_id;
char* tmp_ptr;
// TODO: be annotation-aware
codes_mapping_get_lp_info(sender->gid, grp_name, &mapping_grp_id, NULL,
&mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
mapping_rep_id, mapping_offset, &local_nic_id);
xfer_to_nic_time = g_tw_lookahead + codes_local_latency(sender);
e_new = model_net_method_event_new(local_nic_id, xfer_to_nic_time,
sender, TORUS, (void**)&msg, (void**)&tmp_ptr);
msg->remote_event_size_bytes = message_size;
strcpy(msg->category, category);
msg->sender_svr=sender->gid;
msg->type = T_COLLECTIVE_INIT;
tmp_ptr = (char*)msg;
tmp_ptr += torus_get_msg_sz();
if(remote_event_size > 0)
{
msg->remote_event_size_bytes = remote_event_size;
memcpy(tmp_ptr, remote_event, remote_event_size);
tmp_ptr += remote_event_size;
}
tw_event_send(e_new);
return;
}
void torus_collective_init(nodes_state * s,
tw_lp * lp)
{
// TODO: be annotation-aware somehow
codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM, s->anno, 0);
int num_reps = codes_mapping_get_group_reps(grp_name);
s->node_id = (mapping_rep_id * num_lps) + mapping_offset;
int i;
/* handle collective operations by forming a tree of all the LPs */
/* special condition for root of the tree */
if( s->node_id == 0)
{
s->parent_node_id = -1;
s->is_root = 1;
}
else
{
s->parent_node_id = (s->node_id - ((s->node_id - 1) % TREE_DEGREE)) / TREE_DEGREE;
s->is_root = 0;
}
s->children = (tw_lpid*)malloc(TREE_DEGREE * sizeof(tw_lpid));
/* set the isleaf to zero by default */
s->is_leaf = 1;
s->num_children = 0;
/* calculate the children of the current node. If its a leaf, no need to set children,
only set isleaf and break the loop*/
for( i = 0; i < TREE_DEGREE; i++ )
{
tw_lpid next_child = (TREE_DEGREE * s->node_id) + i + 1;
if(next_child < (num_lps * num_reps))
{
s->num_children++;
s->is_leaf = 0;
s->children[i] = next_child;
}
else
s->children[i] = -1;
}
#if TORUS_COLLECTIVE_DEBUG == 1
printf("\n LP %ld parent node id ", s->node_id);
for( i = 0; i < TREE_DEGREE; i++ )
printf(" child node ID %ld ", s->children[i]);
printf("\n");
if(s->is_leaf)
printf("\n LP %ld is leaf ", s->node_id);
#endif
}
tw_lpid get_next_node(tw_lpid sender_id) {
tw_lpid rtn_id;
char grp_name[MAX_NAME_LENGTH], lp_type_name[MAX_NAME_LENGTH],
annotation[MAX_NAME_LENGTH];
int lp_type_id, grp_id, grp_rep_id, offset; //num_reps;
codes_mapping_get_lp_info(sender_id, grp_name, &grp_id, lp_type_name,
&lp_type_id, annotation, &grp_rep_id, &offset);
// num_reps = codes_mapping_get_group_reps(grp_name);
//We match up node pairs randomly.
// while((dest_rep_id = rand() % num_reps) == grp_rep_id);
if (grp_rep_id != 1) rtn_id = (tw_lpid) -1;
else codes_mapping_get_lp_id(grp_name, lp_type_name, NULL, 1, 2/*the neighbor*/, 0, &rtn_id);
return rtn_id;
}
static void node_collective_init(nodes_state * s,
tw_bf * bf,
nodes_message * msg,
tw_lp * lp)
{
tw_event * e_new;
tw_lpid parent_nic_id;
tw_stime xfer_to_nic_time;
nodes_message * msg_new;
int num_lps;
msg->saved_collective_init_time = s->collective_init_time;
s->collective_init_time = tw_now(lp);
s->origin_svr = msg->sender_svr;
if(s->is_leaf)
{
//printf("\n LP %ld sending message to parent %ld ", s->node_id, s->parent_node_id);
/* get the global LP ID of the parent node */
// TODO: be annotation-aware
codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL,
&mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
NULL, 1);
codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
&parent_nic_id);
/* send a message to the parent that the LP has entered the collective operation */
xfer_to_nic_time = g_tw_lookahead + LEVEL_DELAY;
//e_new = codes_event_new(parent_nic_id, xfer_to_nic_time, lp);
void* m_data;
e_new = model_net_method_event_new(parent_nic_id, xfer_to_nic_time,
lp, TORUS, (void**)&msg_new, (void**)&m_data);
memcpy(msg_new, msg, sizeof(nodes_message));
if (msg->remote_event_size_bytes) {
memcpy(m_data, model_net_method_get_edata(TORUS, msg),
msg->remote_event_size_bytes);
}
msg_new->type = T_COLLECTIVE_FAN_IN;
msg_new->sender_node = s->node_id;
tw_event_send(e_new);
}
return;
}
static void handle_kickoff_event(
svr_state * ns,
tw_bf * b,
svr_msg * m,
tw_lp * lp)
{
char* anno;
tw_lpid local_dest = -1, global_dest = -1;
svr_msg * m_local = malloc(sizeof(svr_msg));
svr_msg * m_remote = malloc(sizeof(svr_msg));
m_local->svr_event_type = LOCAL;
m_local->src = lp->gid;
memcpy(m_remote, m_local, sizeof(svr_msg));
m_remote->svr_event_type = REMOTE;
assert(net_id == DRAGONFLY); /* only supported for dragonfly model right now. */
ns->start_ts = tw_now(lp);
codes_mapping_get_lp_info(lp->gid, group_name, &group_index, lp_type_name, &lp_type_index, anno, &rep_id, &offset);
/* in case of uniform random traffic, send to a random destination. */
if(traffic == UNIFORM)
{
local_dest = tw_rand_integer(lp->rng, 0, num_nodes - 1);
// printf("\n LP %ld sending to %d ", lp->gid, local_dest);
}
else if(traffic == NEAREST_GROUP)
{
local_dest = (rep_id * 2 + offset + num_nodes_per_grp) % num_nodes;
// printf("\n LP %ld sending to %ld num nodes %d ", rep_id * 2 + offset, local_dest, num_nodes);
}
else if(traffic == NEAREST_NEIGHBOR)
{
local_dest = (rep_id * 2 + offset + 2) % num_nodes;
// printf("\n LP %ld sending to %ld num nodes %d ", rep_id * 2 + offset, local_dest, num_nodes);
}
assert(local_dest < num_nodes);
codes_mapping_get_lp_id(group_name, lp_type_name, anno, 1, local_dest / num_servers_per_rep, local_dest % num_servers_per_rep, &global_dest);
ns->msg_sent_count++;
model_net_event(net_id, "test", global_dest, PAYLOAD_SZ, 0.0, sizeof(svr_msg), (const void*)m_remote, sizeof(svr_msg), (const void*)m_local, lp);
issue_event(ns, lp);
return;
}
static void node_collective_fan_in(nodes_state * s,
tw_bf * bf,
nodes_message * msg,
tw_lp * lp)
{
int i;
s->num_fan_nodes++;
// TODO: be annotation-aware
codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL,
&mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
NULL, 1);
tw_event* e_new;
nodes_message * msg_new;
tw_stime xfer_to_nic_time;
bf->c1 = 0;
bf->c2 = 0;
/* if the number of fanned in nodes have completed at the current node then signal the parent */
if((s->num_fan_nodes == s->num_children) && !s->is_root)
{
bf->c1 = 1;
msg->saved_fan_nodes = s->num_fan_nodes-1;
s->num_fan_nodes = 0;
tw_lpid parent_nic_id;
xfer_to_nic_time = g_tw_lookahead + LEVEL_DELAY;
/* get the global LP ID of the parent node */
codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
&parent_nic_id);
/* send a message to the parent that the LP has entered the collective operation */
//e_new = codes_event_new(parent_nic_id, xfer_to_nic_time, lp);
//msg_new = tw_event_data(e_new);
void * m_data;
e_new = model_net_method_event_new(parent_nic_id,
xfer_to_nic_time,
lp, TORUS, (void**)&msg_new, &m_data);
memcpy(msg_new, msg, sizeof(nodes_message));
msg_new->type = T_COLLECTIVE_FAN_IN;
msg_new->sender_node = s->node_id;
if (msg->remote_event_size_bytes) {
memcpy(m_data, model_net_method_get_edata(TORUS, msg),
msg->remote_event_size_bytes);
}
tw_event_send(e_new);
}
/* root node starts off with the fan-out phase */
if(s->is_root && (s->num_fan_nodes == s->num_children))
{
bf->c2 = 1;
msg->saved_fan_nodes = s->num_fan_nodes-1;
s->num_fan_nodes = 0;
send_remote_event(s, bf, msg, lp);
for( i = 0; i < s->num_children; i++ )
{
tw_lpid child_nic_id;
/* Do some computation and fan out immediate child nodes from the collective */
xfer_to_nic_time = g_tw_lookahead + COLLECTIVE_COMPUTATION_DELAY + LEVEL_DELAY + tw_rand_exponential(lp->rng, (double)LEVEL_DELAY/50);
/* get global LP ID of the child node */
codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
s->children[i]/num_lps, (s->children[i] % num_lps),
&child_nic_id);
//e_new = codes_event_new(child_nic_id, xfer_to_nic_time, lp);
//msg_new = tw_event_data(e_new);
void * m_data;
e_new = model_net_method_event_new(child_nic_id,
xfer_to_nic_time,
lp, TORUS, (void**)&msg_new, &m_data);
memcpy(msg_new, msg, sizeof(nodes_message));
if (msg->remote_event_size_bytes) {
memcpy(m_data, model_net_method_get_edata(TORUS, msg),
msg->remote_event_size_bytes);
}
msg_new->type = T_COLLECTIVE_FAN_OUT;
msg_new->sender_node = s->node_id;
tw_event_send(e_new);
}
}
}
/*Initialize the torus model, this initialization part is borrowed from Ning's torus model */
static void torus_init( nodes_state * s,
tw_lp * lp )
{
int i, j;
char anno[MAX_NAME_LENGTH];
codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL, &mapping_type_id, anno, &mapping_rep_id, &mapping_offset);
if (anno[0] == '\0') {
s->anno = NULL;
s->params = &all_params[num_params-1];
}
else {
s->anno = strdup(anno);
int id = configuration_get_annotation_index(anno, anno_map);
s->params = &all_params[id];
}
// shorthand
const torus_param *p = s->params;
s->neighbour_minus_lpID = (int*)malloc(p->n_dims * sizeof(int));
s->neighbour_plus_lpID = (int*)malloc(p->n_dims * sizeof(int));
s->dim_position = (int*)malloc(p->n_dims * sizeof(int));
s->buffer = (int**)malloc(2*p->n_dims * sizeof(int*));
s->next_link_available_time =
(tw_stime**)malloc(2*p->n_dims * sizeof(tw_stime*));
s->next_credit_available_time =
(tw_stime**)malloc(2*p->n_dims * sizeof(tw_stime*));
s->next_flit_generate_time =
(tw_stime**)malloc(2*p->n_dims*sizeof(tw_stime*));
for(i=0; i < 2*p->n_dims; i++)
{
s->buffer[i] = (int*)malloc(p->num_vc * sizeof(int));
s->next_link_available_time[i] =
(tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
s->next_credit_available_time[i] =
(tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
s->next_flit_generate_time[i] =
(tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
}
// calculate my torus coords
to_dim_id(codes_mapping_get_lp_relative_id(lp->gid, 0, 1),
s->params->n_dims, s->params->dim_length, s->dim_position);
/* DEBUG
printf("%lu: my coords:", lp->gid);
for (i = 0; i < p->n_dims; i++)
printf(" %d", s->dim_position[i]);
printf("\n");
*/
int temp_dim_pos[ p->n_dims ];
for ( i = 0; i < p->n_dims; i++ )
temp_dim_pos[ i ] = s->dim_position[ i ];
// calculate minus neighbour's lpID
for ( j = 0; j < p->n_dims; j++ )
{
temp_dim_pos[ j ] = (s->dim_position[ j ] -1 + p->dim_length[ j ]) %
p->dim_length[ j ];
s->neighbour_minus_lpID[j] =
to_flat_id(p->n_dims, p->dim_length, temp_dim_pos);
/* DEBUG
printf(" minus neighbor: flat:%d lpid:%lu\n",
s->neighbour_minus_lpID[j],
codes_mapping_get_lpid_from_relative(s->neighbour_minus_lpID[j],
NULL, LP_CONFIG_NM, s->anno, 1));
*/
temp_dim_pos[ j ] = s->dim_position[ j ];
}
// calculate plus neighbour's lpID
for ( j = 0; j < p->n_dims; j++ )
{
temp_dim_pos[ j ] = ( s->dim_position[ j ] + 1 + p->dim_length[ j ]) %
p->dim_length[ j ];
s->neighbour_plus_lpID[j] =
to_flat_id(p->n_dims, p->dim_length, temp_dim_pos);
/* DEBUG
printf(" plus neighbor: flat:%d lpid:%lu\n",
s->neighbour_plus_lpID[j],
codes_mapping_get_lpid_from_relative(s->neighbour_plus_lpID[j],
NULL, LP_CONFIG_NM, s->anno, 1));
*/
temp_dim_pos[ j ] = s->dim_position[ j ];
}
//printf("\n");
for( j=0; j < 2 * p->n_dims; j++ )
{
for( i = 0; i < p->num_vc; i++ )
{
s->buffer[ j ][ i ] = 0;
s->next_link_available_time[ j ][ i ] = 0.0;
s->next_credit_available_time[j][i] = 0.0;
}
}
// record LP time
s->packet_counter = 0;
torus_collective_init(s, lp);
}