void
tw_gvt_step2(tw_pe *me)
{
tw_stat local_white = 0;
tw_stat total_white = 0;
tw_stime pq_min = DBL_MAX;
tw_stime net_min = DBL_MAX;
tw_stime lvt;
tw_stime gvt;
tw_clock start = tw_clock_read();
if(me->gvt_status != TW_GVT_COMPUTE)
return;
while(1)
{
tw_net_read(me);
// send message counts to create consistent cut
local_white = me->s_nwhite_sent - me->s_nwhite_recv;
all_reduce_cnt++;
if(MPI_Allreduce(
&local_white,
&total_white,
1,
MPI_LONG_LONG,
MPI_SUM,
MPI_COMM_WORLD) != MPI_SUCCESS)
tw_error(TW_LOC, "MPI_Allreduce for GVT failed");
if(total_white == 0)
break;
}
pq_min = tw_pq_minimum(me->pq);
net_min = tw_net_minimum(me);
lvt = me->trans_msg_ts;
if(lvt > pq_min)
lvt = pq_min;
if(lvt > net_min)
lvt = net_min;
all_reduce_cnt++;
if(MPI_Allreduce(
&lvt,
&gvt,
1,
MPI_DOUBLE,
MPI_MIN,
MPI_COMM_WORLD) != MPI_SUCCESS)
tw_error(TW_LOC, "MPI_Allreduce for GVT failed");
gvt = min(gvt, me->GVT_prev);
if(gvt != me->GVT_prev)
{
g_tw_gvt_no_change = 0;
} else
{
g_tw_gvt_no_change++;
if (g_tw_gvt_no_change >= g_tw_gvt_max_no_change) {
tw_error(
TW_LOC,
"GVT computed %d times in a row"
" without changing: GVT = %14.14lf, PREV %14.14lf"
" -- GLOBAL SYNCH -- out of memory!",
g_tw_gvt_no_change, gvt, me->GVT_prev);
}
}
if (me->GVT > gvt)
{
tw_error(TW_LOC, "PE %u GVT decreased %g -> %g",
me->id, me->GVT, gvt);
}
if(gvt / g_tw_ts_end > percent_complete &&
tw_node_eq(&g_tw_mynode, &g_tw_masternode))
{
gvt_print(gvt);
}
me->s_nwhite_sent = 0;
me->s_nwhite_recv = 0;
me->trans_msg_ts = DBL_MAX;
me->GVT_prev = DBL_MAX; // me->GVT;
me->GVT = gvt;
me->gvt_status = TW_GVT_NORMAL;
// printf("gvt before fossil %14.14lf\n", gvt);
gvt_cnt = 0;
// update GVT timing stats
me->stats.s_gvt += tw_clock_read() - start;
// only FC if OPTIMISTIC
if( g_tw_synchronization_protocol == OPTIMISTIC )
{
start = tw_clock_read();
// printf("ROSS fossil collect at gvt %14.14lf\n", gvt);
tw_pe_fossil_collect(me);
fossil_collected = 1;
me->stats.s_fossil_collect += tw_clock_read() - start;
}
g_tw_gvt_done++;
}
void
tw_gvt_step2(tw_pe *me)
{
long long local_white = 0;
long long total_white = 0;
tw_stime pq_min = DBL_MAX;
tw_stime net_min = DBL_MAX;
tw_stime lvt;
tw_stime gvt;
tw_clock net_start;
tw_clock start = tw_clock_read();
if(me->gvt_status != TW_GVT_COMPUTE)
return;
while(1)
{
net_start = tw_clock_read();
tw_net_read(me);
me->stats.s_net_read += tw_clock_read() - net_start;
// send message counts to create consistent cut
local_white = me->s_nwhite_sent - me->s_nwhite_recv;
all_reduce_cnt++;
if(MPI_Allreduce(
&local_white,
&total_white,
1,
MPI_LONG_LONG,
MPI_SUM,
MPI_COMM_ROSS) != MPI_SUCCESS)
tw_error(TW_LOC, "MPI_Allreduce for GVT failed");
if(total_white == 0)
break;
}
pq_min = tw_pq_minimum(me->pq);
net_min = tw_net_minimum(me);
lvt = me->trans_msg_ts;
if(lvt > pq_min)
lvt = pq_min;
if(lvt > net_min)
lvt = net_min;
all_reduce_cnt++;
if(MPI_Allreduce(
&lvt,
&gvt,
1,
MPI_DOUBLE,
MPI_MIN,
MPI_COMM_ROSS) != MPI_SUCCESS)
tw_error(TW_LOC, "MPI_Allreduce for GVT failed");
gvt = ROSS_MIN(gvt, me->GVT_prev);
if(gvt != me->GVT_prev)
{
g_tw_gvt_no_change = 0;
} else
{
g_tw_gvt_no_change++;
if (g_tw_gvt_no_change >= g_tw_gvt_max_no_change) {
tw_error(
TW_LOC,
"GVT computed %d times in a row"
" without changing: GVT = %14.14lf, PREV %14.14lf"
" -- GLOBAL SYNCH -- out of memory!",
g_tw_gvt_no_change, gvt, me->GVT_prev);
}
}
if (me->GVT > gvt)
{
tw_error(TW_LOC, "PE %u GVT decreased %g -> %g",
me->id, me->GVT, gvt);
}
if (gvt / g_tw_ts_end > percent_complete && (g_tw_mynode == g_tw_masternode))
{
gvt_print(gvt);
}
me->s_nwhite_sent = 0;
me->s_nwhite_recv = 0;
me->trans_msg_ts = DBL_MAX;
me->GVT_prev = DBL_MAX; // me->GVT;
me->GVT = gvt;
me->gvt_status = TW_GVT_NORMAL;
gvt_cnt = 0;
// update GVT timing stats
me->stats.s_gvt += tw_clock_read() - start;
// only FC if OPTIMISTIC or REALTIME, do not do for DEBUG MODE
if( g_tw_synchronization_protocol == OPTIMISTIC ||
g_tw_synchronization_protocol == OPTIMISTIC_REALTIME )
{
start = tw_clock_read();
tw_pe_fossil_collect(me);
me->stats.s_fossil_collect += tw_clock_read() - start;
}
// do any necessary instrumentation calls
if ((g_st_engine_stats == GVT_STATS || g_st_engine_stats == ALL_STATS) &&
g_tw_gvt_done % g_st_num_gvt == 0 && gvt <= g_tw_ts_end)
{
#ifdef USE_DAMARIS
if (g_st_damaris_enabled)
{
st_damaris_expose_data(me, gvt, GVT_COL);
st_damaris_end_iteration();
}
else
st_collect_engine_data(me, GVT_COL);
#else
st_collect_engine_data(me, GVT_COL);
#endif
}
#ifdef USE_DAMARIS
// need to make sure damaris_end_iteration is called if GVT instrumentation not turned on
//if (!g_st_stats_enabled && g_st_real_time_samp) //need to make sure if one PE enters this, all do; otherwise deadlock
if (g_st_damaris_enabled && (g_st_engine_stats == RT_STATS || g_st_engine_stats == VT_STATS))
{
st_damaris_end_iteration();
}
#endif
if ((g_st_model_stats == GVT_STATS || g_st_model_stats == ALL_STATS) && g_tw_gvt_done % g_st_num_gvt == 0)
st_collect_model_data(me, (tw_stime)tw_clock_read() / g_tw_clock_rate, GVT_STATS);
st_inst_dump();
// done with instrumentation related stuff
g_tw_gvt_done++;
// reset for the next gvt round -- for use in realtime GVT mode only!!
g_tw_gvt_interval_start_cycles = tw_clock_read();
}