static inline void event_cancel(tw_event * event) {
tw_pe *send_pe = event->src_lp->pe;
tw_peid dest_peid;
if(event->state.owner == TW_net_asend || event->state.owner == TW_pe_sevent_q) {
/* Slowest approach of all; this has to be sent over the
* network to let the dest_pe know it shouldn't have seen
* it in the first place.
*/
tw_net_cancel(event);
send_pe->stats.s_nsend_net_remote--;
if(tw_gvt_inprogress(send_pe)) {
send_pe->trans_msg_ts = ROSS_MIN(send_pe->trans_msg_ts, event->recv_ts);
}
return;
}
dest_peid = event->dest_lp->pe->id;
if (send_pe->id == dest_peid) {
switch (event->state.owner) {
case TW_pe_pq:
/* Currently in our pq and not processed; delete it and
* free the event buffer immediately. No need to wait.
*/
tw_pq_delete_any(send_pe->pq, event);
tw_event_free(send_pe, event);
break;
case TW_pe_event_q:
case TW_kp_pevent_q:
local_cancel(send_pe, event);
if(tw_gvt_inprogress(send_pe)) {
send_pe->trans_msg_ts = ROSS_MIN(send_pe->trans_msg_ts, event->recv_ts);
}
break;
default:
tw_error(TW_LOC, "unknown fast local cancel owner %d", event->state.owner);
}
} else if (send_pe->node == dest_peid) {
/* Slower, but still a local cancel, so put into
* top of dest_pe->cancel_q for final deletion.
*/
local_cancel(event->dest_lp->pe, event);
send_pe->stats.s_nsend_loc_remote--;
if(tw_gvt_inprogress(send_pe)) {
send_pe->trans_msg_ts = ROSS_MIN(send_pe->trans_msg_ts, event->recv_ts);
}
} else {
tw_error(TW_LOC, "Should be remote cancel!");
}
}
static int
recv_begin(tw_pe *me)
{
MPI_Status status;
tw_event *e = NULL;
int flag = 0;
int changed = 0;
while (posted_recvs.cur < read_buffer)
{
unsigned id = posted_recvs.cur;
if(!(e = tw_event_grab(me)))
{
if(tw_gvt_inprogress(me))
tw_error(TW_LOC, "Out of events in GVT! Consider increasing --extramem");
return changed;
}
#if ROSS_MEMORY
if( MPI_Irecv(posted_recvs.buffers[id],
EVENT_SIZE(e),
MPI_BYTE,
MPI_ANY_SOURCE,
EVENT_TAG,
MPI_COMM_ROSS,
&posted_recvs.req_list[id]) != MPI_SUCCESS)
#else
if( MPI_Irecv(e,
(int)EVENT_SIZE(e),
MPI_BYTE,
MPI_ANY_SOURCE,
EVENT_TAG,
MPI_COMM_ROSS,
&posted_recvs.req_list[id]) != MPI_SUCCESS)
#endif
{
tw_event_free(me, e);
return changed;
}
posted_recvs.event_list[id] = e;
posted_recvs.cur++;
changed = 1;
}
return changed;
}
void tw_event_send(tw_event * event) {
tw_lp *src_lp = event->src_lp;
tw_pe *send_pe = src_lp->pe;
tw_pe *dest_pe = NULL;
tw_peid dest_peid = -1;
tw_stime recv_ts = event->recv_ts;
if (event == send_pe->abort_event) {
if (recv_ts < g_tw_ts_end) {
send_pe->cev_abort = 1;
}
return;
}
//Trap lookahead violations in debug mode
//Note that compiling with the -DNDEBUG flag will turn this off!
if (g_tw_synchronization_protocol == CONSERVATIVE) {
if (recv_ts - tw_now(src_lp) < g_tw_lookahead) {
tw_error(TW_LOC, "Lookahead violation: decrease g_tw_lookahead");
}
}
if (event->out_msgs) {
tw_error(TW_LOC, "It is an error to send an event with pre-loaded output message.");
}
link_causality(event, send_pe->cur_event);
// call LP remote mapping function to get dest_pe
dest_peid = (*src_lp->type->map) ((tw_lpid) event->dest_lp);
if (dest_peid == g_tw_mynode) {
event->dest_lp = tw_getlocal_lp((tw_lpid) event->dest_lp);
dest_pe = event->dest_lp->pe;
if (send_pe == dest_pe && event->dest_lp->kp->last_time <= recv_ts) {
/* Fast case, we are sending to our own PE and there is
* no rollback caused by this send. We cannot have any
* transient messages on local sends so we can return.
*/
tw_pq_enqueue(send_pe->pq, event);
return;
} else {
/* Slower, but still local send, so put into top of
* dest_pe->event_q.
*/
event->state.owner = TW_pe_event_q;
tw_eventq_push(&dest_pe->event_q, event);
if(send_pe != dest_pe) {
send_pe->stats.s_nsend_loc_remote++;
}
}
} else {
/* Slowest approach of all; this is not a local event.
* We need to send it over the network to the other PE
* for processing.
*/
send_pe->stats.s_nsend_net_remote++;
event->state.owner = TW_net_asend;
tw_net_send(event);
}
if(tw_gvt_inprogress(send_pe)) {
send_pe->trans_msg_ts = ROSS_MIN(send_pe->trans_msg_ts, recv_ts);
}
}
/**
* @brief Determines how to handle the newly received event.
*
* @param[in] me pointer to PE
* @param[in] e pointer to event that we just received
* @param[in] buffer not currently used
*/
static void
recv_finish(tw_pe *me, tw_event *e, char * buffer)
{
(void) buffer;
tw_pe *dest_pe;
tw_clock start;
me->stats.s_nread_network++;
me->s_nwhite_recv++;
// printf("recv_finish: remote event [cancel %u] FROM: LP %lu, PE %lu, TO: LP %lu, PE %lu at TS %lf \n",
// e->state.cancel_q, (tw_lpid)e->src_lp, e->send_pe, (tw_lpid)e->dest_lp, me->id, e->recv_ts);
e->dest_lp = tw_getlocal_lp((tw_lpid) e->dest_lp);
dest_pe = e->dest_lp->pe;
// instrumentation
e->dest_lp->kp->kp_stats->s_nread_network++;
e->dest_lp->lp_stats->s_nread_network++;
if(e->send_pe > tw_nnodes()-1)
tw_error(TW_LOC, "bad sendpe_id: %d", e->send_pe);
e->cancel_next = NULL;
e->caused_by_me = NULL;
e->cause_next = NULL;
if(e->recv_ts < me->GVT)
tw_error(TW_LOC, "%d: Received straggler from %d: %lf (%d)",
me->id, e->send_pe, e->recv_ts, e->state.cancel_q);
if(tw_gvt_inprogress(me))
me->trans_msg_ts = ROSS_MIN(me->trans_msg_ts, e->recv_ts);
// if cancel event, retrieve and flush
// else, store in hash table
if(e->state.cancel_q)
{
tw_event *cancel = tw_hash_remove(me->hash_t, e, e->send_pe);
// NOTE: it is possible to cancel the event we
// are currently processing at this PE since this
// MPI module lets me read cancel events during
// event sends over the network.
cancel->state.cancel_q = 1;
cancel->state.remote = 0;
cancel->cancel_next = dest_pe->cancel_q;
dest_pe->cancel_q = cancel;
tw_event_free(me, e);
return;
}
if (g_tw_synchronization_protocol == OPTIMISTIC ||
g_tw_synchronization_protocol == OPTIMISTIC_DEBUG ||
g_tw_synchronization_protocol == OPTIMISTIC_REALTIME ) {
tw_hash_insert(me->hash_t, e, e->send_pe);
e->state.remote = 1;
}
/* NOTE: the final check in the if conditional below was added to make sure
* that we do not execute the fast case unless the cancellation queue is
* empty on the destination PE. Otherwise we need to invoke the normal
* scheduling routines to make sure that a forward event doesn't bypass a
* cancellation event with an earlier timestamp. This is helpful for
* stateful models that produce incorrect results when presented with
* duplicate messages with no rollback between them.
*/
if(me == dest_pe && e->dest_lp->kp->last_time <= e->recv_ts && !dest_pe->cancel_q) {
/* Fast case, we are sending to our own PE and
* there is no rollback caused by this send.
*/
start = tw_clock_read();
tw_pq_enqueue(dest_pe->pq, e);
dest_pe->stats.s_pq += tw_clock_read() - start;
return;
}
if (me->id == dest_pe->id) {
/* Slower, but still local send, so put into top
* of dest_pe->event_q.
*/
e->state.owner = TW_pe_event_q;
tw_eventq_push(&dest_pe->event_q, e);
return;
}
/* Never should happen; MPI should have gotten the
* message to the correct node without needing us
* to redirect the message there for it. This is
* probably a serious bug with the event headers
* not being formatted right.
*/
tw_error(
TW_LOC,
"Event recived by PE %u but meant for PE %u",
me->id,
dest_pe->id);
}
static int
recv_begin(tw_pe *me)
{
MPI_Status status;
tw_event *e = NULL;
int flag = 0;
int changed = 0;
while (posted_recvs.cur < read_buffer)
{
unsigned id = posted_recvs.cur;
MPI_Iprobe(MPI_ANY_SOURCE,
MPI_ANY_TAG,
MPI_COMM_WORLD,
&flag,
&status);
if(flag)
{
if(!(e = tw_event_grab(me)))
{
if(tw_gvt_inprogress(me))
tw_error(TW_LOC, "out of events in GVT!");
break;
}
} else
{
return changed;
}
#if ROSS_MEMORY
if(!flag ||
MPI_Irecv(posted_recvs.buffers[id],
EVENT_SIZE(e),
MPI_BYTE,
MPI_ANY_SOURCE,
EVENT_TAG,
MPI_COMM_WORLD,
&posted_recvs.req_list[id]) != MPI_SUCCESS)
#else
if(!flag ||
MPI_Irecv(e,
(int)EVENT_SIZE(e),
MPI_BYTE,
MPI_ANY_SOURCE,
EVENT_TAG,
MPI_COMM_WORLD,
&posted_recvs.req_list[id]) != MPI_SUCCESS)
#endif
{
tw_event_free(me, e);
return changed;
}
posted_recvs.event_list[id] = e;
posted_recvs.cur++;
changed = 1;
}
return changed;
}
