Runtime option for limiting optimistic lookahead (ROSS-org#107)

In irregular simulations, runtime crashes are caused due to memory overflow.
This often happens because some LPs over-optimisitcally explore events in
virtual time and are repeatedly rolled back. One user-defined way of avoiding
these crashes is to restrict the optimistic lookahead in terms of virtual time.

Option added: max-opt-lookahead
Unit: virtual time

If max-opt-lookahead is set, a PE repeatedly computes GVT till
(LVT - GVT) > max-opt-lookahead.

Change-Id: I0f81b75628125e1fd2e71eddaa4116025c00a1c6

core/gvt/mpi_allreduce.c

@@ -65,7 +65,7 @@ void
 tw_gvt_step1(tw_pe *me)
 {
 	if(me->gvt_status == TW_GVT_COMPUTE ||
-		++gvt_cnt < g_tw_gvt_interval)
+		(++gvt_cnt < g_tw_gvt_interval && (tw_pq_minimum(me->pq) - me->GVT < g_tw_max_opt_lookahead)))
 		return;
 
 	me->gvt_status = TW_GVT_COMPUTE;
@@ -77,7 +77,8 @@ tw_gvt_step1_realtime(tw_pe *me)
   unsigned long long current_rt;
 
   if( (me->gvt_status == TW_GVT_COMPUTE) ||
-      ( (current_rt = tw_clock_read()) - g_tw_gvt_interval_start_cycles < g_tw_gvt_realtime_interval))
+      ( ((current_rt = tw_clock_read()) - g_tw_gvt_interval_start_cycles < g_tw_gvt_realtime_interval)
+          && (tw_pq_minimum(me->pq) - me->GVT < g_tw_max_opt_lookahead)))
     {
       /* if( me->node == 0 ) */
       /* 	{ */

core/ross-extern.h

@@ -27,6 +27,7 @@ extern tw_lpid		g_tw_rng_default;
 extern tw_seed		g_tw_rng_seed;
 extern unsigned int	g_tw_mblock;
 extern unsigned int g_tw_gvt_interval;
+extern unsigned long long g_tw_max_opt_lookahead;
 extern unsigned long long g_tw_gvt_realtime_interval;
 extern unsigned long long g_tw_gvt_interval_start_cycles;
 extern tw_stime		g_tw_ts_end;

core/ross-global.c

@@ -68,7 +68,7 @@ tw_stime g_tw_min_detected_offset=DBL_MAX;
 	 */
 unsigned int g_tw_mblock = 16;
 unsigned int g_tw_gvt_interval = 16;
-
+unsigned long long g_tw_max_opt_lookahead = ULLONG_MAX;
 unsigned long long g_tw_gvt_realtime_interval; // calculated at runtime
 unsigned long long g_tw_gvt_interval_start_cycles = 0;
 

core/tw-setup.c

@@ -20,6 +20,7 @@ static const tw_optdef kernel_options[] = {
     TWOPT_UINT("extramem", g_tw_events_per_pe_extra, "Number of extra events allocated per PE."),
     TWOPT_UINT("buddy-size", g_tw_buddy_alloc, "delta encoding buddy system allocation (2^X)"),
     TWOPT_UINT("lz4-knob", g_tw_lz4_knob, "LZ4 acceleration factor (higher = faster)"),
+    TWOPT_ULONGLONG("max-opt-lookahead", g_tw_max_opt_lookahead, "Optimistic simulation: maximum lookahead allowed in virtual clock time"),
 #ifdef AVL_TREE
     TWOPT_UINT("avl-size", g_tw_avl_node_count, "AVL Tree contains 2^avl-size nodes"),
 #endif