Add support for scheduler utilization balancing

For more information see documentation of the new command line argument +sub

3 files changed, 497 insertions, 57 deletions

diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c
index 1af80dd04b..1d4f617746 100644
--- a/erts/emulator/beam/erl_init.c
+++ b/erts/emulator/beam/erl_init.c
@@ -537,6 +537,12 @@ void erts_usage(void)
     erts_fprintf(stderr, "            see the erl(1) documentation for more info.\n");
     erts_fprintf(stderr, "-sct cput   set cpu topology,\n");
     erts_fprintf(stderr, "            see the erl(1) documentation for more info.\n");
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT_OPT
+    erts_fprintf(stderr, "-sub bool   enable/disable scheduler utilization balancing,\n");
+#else
+    erts_fprintf(stderr, "-sub false  disable scheduler utilization balancing,\n");
+#endif
+    erts_fprintf(stderr, "            see the erl(1) documentation for more info.\n");
     erts_fprintf(stderr, "-sws val    set scheduler wakeup strategy, valid values are:\n");
     erts_fprintf(stderr, "            default|legacy.\n");
     erts_fprintf(stderr, "-swct val   set scheduler wake cleanup threshold, valid values are:\n");
@@ -1512,6 +1518,26 @@ erl_start(int argc, char **argv)
 		    erts_usage();
 		}
 	    }
+	    else if (has_prefix("ub", sub_param)) {
+		arg = get_arg(sub_param+2, argv[i+1], &i);
+		if (sys_strcmp("true", arg) == 0) {
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT_OPT
+		    erts_sched_balance_util = 1;
+#else
+		    erts_fprintf(stderr,
+				 "scheduler utilization balancing not "
+				 "supported on this system\n");
+		    erts_usage();
+#endif
+		}
+		else if (sys_strcmp("false", arg) == 0)
+		    erts_sched_balance_util = 0;
+		else {
+		    erts_fprintf(stderr, "bad scheduler utilization balancing "
+				 " value '%s'\n", arg);
+		    erts_usage();
+		}
+	    }
 	    else if (has_prefix("wct", sub_param)) {
 		arg = get_arg(sub_param+3, argv[i+1], &i);
 		if (erts_sched_set_wake_cleanup_threshold(arg) != 0) {
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 21fd8dd50a..6143c5fa3a 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -144,6 +144,7 @@ extern BeamInstr beam_exit[];
 extern BeamInstr beam_continue_exit[];
 
 int erts_sched_compact_load;
+int erts_sched_balance_util = 0;
 Uint erts_no_schedulers;
 
 #define ERTS_THR_PRGR_LATER_CLEANUP_OP_THRESHOLD_VERY_LAZY		(4*1024*1024)
@@ -608,6 +609,7 @@ erts_late_init_process(void)
 static void
 init_sched_wall_time(ErtsSchedWallTime *swtp)
 {
+    swtp->need = erts_sched_balance_util;
     swtp->enabled = 0;
     swtp->start = 0;
     swtp->working.total = 0;
@@ -630,27 +632,253 @@ sched_wall_time_ts(void)
 #endif
 }
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+
+#ifdef ARCH_64
+
+static ERTS_INLINE Uint64
+aschedtime_read(ErtsAtomicSchedTime *var)
+{
+    return (Uint64) erts_atomic_read_nob((erts_atomic_t *) var);
+}
+
+static ERTS_INLINE void
+aschedtime_set(ErtsAtomicSchedTime *var, Uint64 val)
+{
+    erts_atomic_set_nob((erts_atomic_t *) var, (erts_aint_t) val);
+}
+
+static ERTS_INLINE void
+aschedtime_init(ErtsAtomicSchedTime *var)
+{
+    erts_atomic_init_nob((erts_atomic_t *) var, (erts_aint_t) 0);
+}
+
+#elif defined(ARCH_32)
+
+static ERTS_INLINE Uint64
+aschedtime_read(ErtsAtomicSchedTime *var)
+{
+    erts_dw_aint_t dw;
+    erts_dw_atomic_read_nob((erts_dw_atomic_t *) var, &dw);
+#ifdef ETHR_SU_DW_NAINT_T__
+    return (Uint64) dw.dw_sint;
+#else
+    {
+	Uint64 res;
+	res = (Uint64) ((Uint32) dw.sint[ERTS_DW_AINT_HIGH_WORD]);
+	res <<= 32;
+	res |= (Uint64) ((Uint32) dw.sint[ERTS_DW_AINT_LOW_WORD]);
+	return res;
+    }
+#endif    
+}
+
+static ERTS_INLINE void
+aschedtime_set(ErtsAtomicSchedTime *var, Uint64 val)
+{
+    erts_dw_aint_t dw;
+#ifdef ETHR_SU_DW_NAINT_T__
+    dw.dw_sint = (ETHR_SU_DW_NAINT_T__) val;
+#else
+    dw.sint[ERTS_DW_AINT_LOW_WORD] = (erts_aint_t) (val & 0xffffffff);
+    dw.sint[ERTS_DW_AINT_HIGH_WORD] = (erts_aint_t) ((val >> 32) & 0xffffffff);
+#endif
+    erts_dw_atomic_set_nob((erts_dw_atomic_t *) var, &dw);
+}
+
+static ERTS_INLINE void
+aschedtime_init(ErtsAtomicSchedTime *var)
+{
+    erts_dw_aint_t dw;
+    dw.sint[ERTS_DW_AINT_LOW_WORD] = (erts_aint_t) 0;
+    dw.sint[ERTS_DW_AINT_HIGH_WORD] = (erts_aint_t) 0;
+    erts_dw_atomic_init_nob((erts_dw_atomic_t *) var, &dw);
+}
+
+#else
+#  error :-/
+#endif
+
+#define ERTS_GET_AVG_MAX_UNLOCKED_TRY 50
+#define ERTS_SCHED_AVG_UTIL_WRITE_MARKER (~((Uint64) 0))
+
+/* Intervals in nanoseconds */
+#define ERTS_SCHED_UTIL_SHORT_INTERVAL ((Uint64) 1*1000*1000*1000)
+#define ERTS_SCHED_UTIL_LONG_INTERVAL ((Uint64) 10*1000*1000*1000)
+
+
+#define ERTS_SCHED_UTIL_IGNORE_IMBALANCE_DIFF 5000 /* ppm */
+
+static ERTS_INLINE Uint64
+calc_sched_worktime(int is_working, Uint64 now, Uint64 last,
+		    Uint64 interval, Uint64 old_worktime)
+{
+    Uint64 worktime;
+    Uint64 new;
+
+    if (now <= last)
+	return old_worktime;
+
+    new = now - last;
+
+    if (new >= interval)
+	return is_working ? interval : (Uint64) 0;
+
+
+    /*
+     * Division by 1000 in order to avoid
+     * overflow. If changed update assertions
+     * in init_runq_sched_util().
+     */
+    worktime = old_worktime;
+    worktime *= (interval - new)/1000;
+    worktime /= (interval/1000);
+    if (is_working)
+	worktime += new;
+
+    ASSERT(0 <= worktime && worktime <= interval);
+
+    return worktime;
+}
+
+static ERTS_INLINE void
+update_avg_sched_util(ErtsSchedulerData *esdp, Uint64 now, int is_working)
+{
+    ErtsRunQueue *rq;
+    int worked;
+    Uint64 swt, lwt, last;
+
+    rq = esdp->run_queue;
+    last = aschedtime_read(&rq->sched_util.last);
+
+    if (now <= last) {
+	ASSERT(last == ERTS_SCHED_AVG_UTIL_WRITE_MARKER);
+	return;
+    }
+
+    ASSERT(now >= last);
+
+    worked = rq->sched_util.is_working;
+
+    swt = calc_sched_worktime(worked, now, last, ERTS_SCHED_UTIL_SHORT_INTERVAL,
+			      rq->sched_util.worktime.short_interval);
+    lwt = calc_sched_worktime(worked, now, last, ERTS_SCHED_UTIL_LONG_INTERVAL,
+			      rq->sched_util.worktime.long_interval);
+
+    aschedtime_set(&rq->sched_util.last, ERTS_SCHED_AVG_UTIL_WRITE_MARKER);
+    ERTS_THR_WRITE_MEMORY_BARRIER;
+    rq->sched_util.is_working = is_working;
+    rq->sched_util.worktime.short_interval = swt;
+    rq->sched_util.worktime.long_interval = lwt;
+    ERTS_THR_WRITE_MEMORY_BARRIER;
+    aschedtime_set(&rq->sched_util.last, now);
+}
+
+int
+erts_get_sched_util(ErtsRunQueue *rq, int initially_locked, int short_interval)
+{
+    /* Average scheduler utilization in ppm */
+    int util, is_working, try = 0, locked = initially_locked;
+    Uint64 worktime, old_worktime, now, last, interval, *old_worktimep;
+
+    if (short_interval) {
+	old_worktimep = &rq->sched_util.worktime.short_interval;
+	interval = ERTS_SCHED_UTIL_SHORT_INTERVAL;
+    }
+    else {
+	old_worktimep = &rq->sched_util.worktime.long_interval;
+	interval = ERTS_SCHED_UTIL_LONG_INTERVAL;
+    }
+
+    while (1) {
+	Uint64 chk_last;
+	last = aschedtime_read(&rq->sched_util.last);
+	ERTS_THR_READ_MEMORY_BARRIER;
+	is_working = rq->sched_util.is_working;
+	old_worktime = *old_worktimep;
+	ERTS_THR_READ_MEMORY_BARRIER;
+	chk_last = aschedtime_read(&rq->sched_util.last);
+	if (chk_last == last)
+	    break;
+	if (!locked) {
+	    if (++try >= ERTS_GET_AVG_MAX_UNLOCKED_TRY) {
+		/* Writer will eventually block on runq-lock */
+		erts_smp_runq_lock(rq);
+		locked = 1;
+	    }
+	}
+    }
+
+    if (!initially_locked && locked)
+	erts_smp_runq_unlock(rq);
+
+    now = sched_wall_time_ts();
+    worktime = calc_sched_worktime(is_working, now, last, interval, old_worktime);
+
+    util = (int) ((worktime * 1000000)/interval);
+
+    ASSERT(0 <= util && util <= 1000000);
+
+    return util;
+}
+
+static void
+init_runq_sched_util(ErtsRunQueueSchedUtil *rqsu, int enabled)
+{
+    aschedtime_init(&rqsu->last);
+    if (!enabled)
+	aschedtime_set(&rqsu->last, ERTS_SCHED_AVG_UTIL_WRITE_MARKER);
+    rqsu->is_working = 0;
+    rqsu->worktime.short_interval = (Uint64) 0;
+    rqsu->worktime.long_interval = (Uint64) 0;
+
+#ifdef DEBUG
+    {
+	Uint64 intrvl;
+	/*
+	 * If one of these asserts fail we may have
+	 * overflow in calc_sched_worktime(). Which
+	 * have to be fixed either by shrinking
+	 * interval size, or fix calculation of
+	 * worktime in calc_sched_worktime().
+	 */
+	intrvl = ERTS_SCHED_UTIL_SHORT_INTERVAL;
+	ASSERT(intrvl*(intrvl/1000) > intrvl);
+	intrvl = ERTS_SCHED_UTIL_LONG_INTERVAL;
+	ASSERT(intrvl*(intrvl/1000) > intrvl);
+    }
+#endif
+}
+
+#endif /* ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT */
+
 static ERTS_INLINE void
 sched_wall_time_change(ErtsSchedulerData *esdp, int working)
 {
-    if (esdp->sched_wall_time.enabled) {
+    if (esdp->sched_wall_time.need) {
 	Uint64 ts = sched_wall_time_ts();
-	if (working) {
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+	update_avg_sched_util(esdp, ts, working);
+#endif
+	if (esdp->sched_wall_time.enabled) {
+	    if (working) {
 #ifdef DEBUG
-	    ASSERT(!esdp->sched_wall_time.working.currently);
-	    esdp->sched_wall_time.working.currently = 1;
+		ASSERT(!esdp->sched_wall_time.working.currently);
+		esdp->sched_wall_time.working.currently = 1;
 #endif
-	    ts -= esdp->sched_wall_time.start;
-	    esdp->sched_wall_time.working.start = ts;
-	}
-	else {
+		ts -= esdp->sched_wall_time.start;
+		esdp->sched_wall_time.working.start = ts;
+	    }
+	    else {
 #ifdef DEBUG
-	    ASSERT(esdp->sched_wall_time.working.currently);
-	    esdp->sched_wall_time.working.currently = 0;
+		ASSERT(esdp->sched_wall_time.working.currently);
+		esdp->sched_wall_time.working.currently = 0;
 #endif
-	    ts -= esdp->sched_wall_time.start;
-	    ts -= esdp->sched_wall_time.working.start;
-	    esdp->sched_wall_time.working.total += ts;
+		ts -= esdp->sched_wall_time.start;
+		ts -= esdp->sched_wall_time.working.start;
+		esdp->sched_wall_time.working.total += ts;
+	    }
 	}
     }
 }
@@ -705,10 +933,13 @@ reply_sched_wall_time(void *vswtrp)
     ASSERT(esdp);
     
     if (swtrp->set) {
-	if (!swtrp->enable && esdp->sched_wall_time.enabled)
+	if (!swtrp->enable && esdp->sched_wall_time.enabled) {
+	    esdp->sched_wall_time.need = erts_sched_balance_util;
 	    esdp->sched_wall_time.enabled = 0;
+	}
 	else if (swtrp->enable && !esdp->sched_wall_time.enabled) {
 	    Uint64 ts = sched_wall_time_ts();
+	    esdp->sched_wall_time.need = 1;
 	    esdp->sched_wall_time.enabled = 1;
 	    esdp->sched_wall_time.start = ts;
 	    esdp->sched_wall_time.working.total = 0;
@@ -2084,9 +2315,8 @@ ongoing_multi_scheduling_block(void)
 }
 
 static ERTS_INLINE void
-empty_runq(ErtsRunQueue *rq)
+empty_runq_aux(ErtsRunQueue *rq, Uint32 old_flags)
 {
-    Uint32 old_flags = ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_NONEMPTY|ERTS_RUNQ_FLG_PROTECTED);
     if (old_flags & ERTS_RUNQ_FLG_NONEMPTY) {
 #ifdef DEBUG
 	erts_aint32_t empty = erts_smp_atomic32_read_nob(&no_empty_run_queues);
@@ -2107,6 +2337,23 @@ empty_runq(ErtsRunQueue *rq)
 }
 
 static ERTS_INLINE void
+empty_runq(ErtsRunQueue *rq)
+{
+    Uint32 old_flags = ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_NONEMPTY|ERTS_RUNQ_FLG_PROTECTED);
+    empty_runq_aux(rq, old_flags);
+}
+
+static ERTS_INLINE Uint32
+empty_protected_runq(ErtsRunQueue *rq)
+{
+    Uint32 old_flags = ERTS_RUNQ_FLGS_BSET(rq,
+					   ERTS_RUNQ_FLG_NONEMPTY|ERTS_RUNQ_FLG_PROTECTED,
+					   ERTS_RUNQ_FLG_PROTECTED);
+    empty_runq_aux(rq, old_flags);
+    return old_flags;
+}
+
+static ERTS_INLINE void
 non_empty_runq(ErtsRunQueue *rq)
 {
     Uint32 old_flags = ERTS_RUNQ_FLGS_SET(rq, ERTS_RUNQ_FLG_NONEMPTY);
@@ -2130,6 +2377,18 @@ non_empty_runq(ErtsRunQueue *rq)
     }
 }
 
+void
+erts_empty_runq(ErtsRunQueue *rq)
+{
+    empty_runq(rq);
+}
+
+void
+erts_non_empty_runq(ErtsRunQueue *rq)
+{
+    non_empty_runq(rq);
+}
+
 static erts_aint32_t
 sched_prep_spin_wait(ErtsSchedulerSleepInfo *ssi)
 {
@@ -2632,7 +2891,7 @@ ssi_flags_set_wake(ErtsSchedulerSleepInfo *ssi)
 }
 
 static void
-wake_scheduler(ErtsRunQueue *rq, int incq)
+wake_scheduler(ErtsRunQueue *rq)
 {
     ErtsSchedulerSleepInfo *ssi;
     erts_aint32_t flgs;
@@ -2651,9 +2910,6 @@ wake_scheduler(ErtsRunQueue *rq, int incq)
 
     flgs = ssi_flags_set_wake(ssi);
     erts_sched_finish_poke(ssi, flgs);
-
-    if (incq && (flgs & ERTS_SSI_FLG_WAITING))
-	non_empty_runq(rq);
 }
 
 #define ERTS_NO_USED_RUNQS_SHIFT 16
@@ -2744,7 +3000,7 @@ chk_wake_sched(ErtsRunQueue *crq, int ix, int activate)
 	    if (try_inc_no_active_runqs(ix+1))
 		(void) ERTS_RUNQ_FLGS_UNSET(wrq, ERTS_RUNQ_FLG_INACTIVE);
 	}
-	wake_scheduler(wrq, 0);
+	wake_scheduler(wrq);
 	return 1;
     }
     return 0;
@@ -2792,7 +3048,7 @@ smp_notify_inc_runq(ErtsRunQueue *runq)
 {
 #ifdef ERTS_SMP
     if (runq)
-	wake_scheduler(runq, 1);
+	wake_scheduler(runq);
 #endif
 }
 
@@ -2810,7 +3066,7 @@ erts_sched_notify_check_cpu_bind(void)
     for (ix = 0; ix < erts_no_run_queues; ix++) {
 	ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
 	(void) ERTS_RUNQ_FLGS_SET(rq, ERTS_RUNQ_FLG_CHK_CPU_BIND);
-	wake_scheduler(rq, 0);
+	wake_scheduler(rq);
     }
 #else
     erts_sched_check_cpu_bind(erts_get_scheduler_data());
@@ -2938,6 +3194,11 @@ check_immigration_need(ErtsRunQueue *c_rq, ErtsMigrationPath *mp, int prio)
     if (!f_rq)
 	return NULL;
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    if (mp->sched_util)
+	return NULL;
+#endif
+
     f_rq_flags = ERTS_RUNQ_FLGS_GET(f_rq);
     if (f_rq_flags & ERTS_RUNQ_FLG_PROTECTED)
 	return NULL;
@@ -3077,7 +3338,7 @@ suspend_run_queue(ErtsRunQueue *rq)
 				   ERTS_SSI_FLG_SUSPENDED);
     (void) ERTS_RUNQ_FLGS_SET(rq, ERTS_RUNQ_FLG_SUSPENDED);
 
-    wake_scheduler(rq, 0);
+    wake_scheduler(rq);
 }
 
 static void scheduler_ix_resume_wake(Uint ix);
@@ -3169,6 +3430,9 @@ evacuate_run_queue(ErtsRunQueue *rq,
 	    to_rq->misc.start = start;
 
 	to_rq->misc.end = end;
+
+	non_empty_runq(to_rq);
+
 	erts_smp_runq_unlock(to_rq);
 	smp_notify_inc_runq(to_rq);
 	erts_smp_runq_lock(to_rq);
@@ -3381,7 +3645,7 @@ try_steal_task(ErtsRunQueue *rq)
     Uint32 flags;
 
     /* Protect jobs we steal from getting stolen from us... */
-    flags = ERTS_RUNQ_FLGS_SET(rq, ERTS_RUNQ_FLG_PROTECTED);
+    flags = empty_protected_runq(rq);
     if (flags & ERTS_RUNQ_FLG_SUSPENDED)
 	return 0; /* go suspend instead... */
 
@@ -3460,6 +3724,9 @@ typedef struct {
     int full_reds_history_change;
     int oowc;
     int max_len;
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    int sched_util;
+#endif
 } ErtsRunQueueBalance;
 static ErtsRunQueueBalance *run_queue_info;
 
@@ -3623,6 +3890,9 @@ check_balance(ErtsRunQueue *c_rq)
     Sint64 scheds_reds, full_scheds_reds;
     int forced, active, current_active, oowc, half_full_scheds, full_scheds,
 	mmax_len, blnc_no_rqs, qix, pix, freds_hist_ix;
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    int sched_util_balancing;
+#endif
 
     if (erts_smp_atomic32_xchg_nob(&balance_info.checking_balance, 1)) {
 	c_rq->check_balance_reds = INT_MAX;
@@ -3678,6 +3948,10 @@ check_balance(ErtsRunQueue *c_rq)
 	return;
     }
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    sched_util_balancing = 0;
+#endif
+
     freds_hist_ix = balance_info.full_reds_history_index;
     balance_info.full_reds_history_index++;
     if (balance_info.full_reds_history_index >= ERTS_FULL_REDS_HISTORY_SIZE)
@@ -3708,7 +3982,12 @@ check_balance(ErtsRunQueue *c_rq)
 	run_queue_info[qix].oowc = rq->out_of_work_count;
 	run_queue_info[qix].max_len = rq->max_len;
 	rq->check_balance_reds = INT_MAX;
-	
+
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+	if (erts_sched_balance_util)
+	    run_queue_info[qix].sched_util = erts_get_sched_util(rq, 1, 0);
+#endif
+
 	erts_smp_runq_unlock(rq);
     }
 
@@ -3778,8 +4057,38 @@ check_balance(ErtsRunQueue *c_rq)
 	    mmax_len = run_queue_info[qix].max_len;
     }
 
-    if (!erts_sched_compact_load)
+    if (!erts_sched_compact_load) {
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+	if (erts_sched_balance_util && full_scheds < blnc_no_rqs) {
+	    int avg_util = 0;
+
+	    for (qix = 0; qix < blnc_no_rqs; qix++)
+		avg_util += run_queue_info[qix].sched_util;
+
+	    avg_util /= blnc_no_rqs; /* in ppm */
+
+	    sched_util_balancing = 1;
+	    /*
+	     * In order to avoid renaming a large amount of fields
+	     * we write utilization values instead of lenght values
+	     * in the 'max_len' and 'migration_limit' fields...
+	     */
+	    for (qix = 0; qix < blnc_no_rqs; qix++) {
+		run_queue_info[qix].flags = 0; /* Reset for later use... */
+		for (pix = 0; pix < ERTS_NO_PRIO_LEVELS; pix++) {
+		    run_queue_info[qix].prio[pix].emigrate_to = -1;
+		    run_queue_info[qix].prio[pix].immigrate_from = -1;
+		    run_queue_info[qix].prio[pix].avail = 100;
+		    run_queue_info[qix].prio[pix].max_len = run_queue_info[qix].sched_util;
+		    run_queue_info[qix].prio[pix].migration_limit = avg_util;
+		}
+	    }
+	    active = blnc_no_rqs;
+	    goto setup_migration_paths;
+	}
+#endif
 	goto all_active;
+    }
 
     if (!forced && half_full_scheds != blnc_no_rqs) {
 	int min = 1;
@@ -3896,15 +4205,30 @@ check_balance(ErtsRunQueue *c_rq)
 	    }
 	}
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    setup_migration_paths:
+#endif
+
 	/* Setup migration paths for all priorities */
 	for (pix = 0; pix < ERTS_NO_PRIO_LEVELS; pix++) {
 	    int low = 0, high = 0;
 	    for (qix = 0; qix < blnc_no_rqs; qix++) {
 		int len_diff = run_queue_info[qix].prio[pix].max_len;
 		len_diff -= run_queue_info[qix].prio[pix].migration_limit;
+
 #ifdef DBG_PRINT
 if (pix == 2) erts_fprintf(stderr, "%d ", len_diff);
 #endif
+
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+		if (sched_util_balancing
+		    && -ERTS_SCHED_UTIL_IGNORE_IMBALANCE_DIFF <= len_diff
+		    && len_diff <= ERTS_SCHED_UTIL_IGNORE_IMBALANCE_DIFF) {
+		    /* ignore minor imbalance */
+		    len_diff = 0;
+		}
+#endif
+		    
 		run_queue_compare[qix].qix = qix;
 		run_queue_compare[qix].len = len_diff;
 		if (len_diff != 0) {
@@ -4031,6 +4355,9 @@ erts_fprintf(stderr, "--------------------------------\n");
 	Uint32 flags = run_queue_info[qix].flags;
 	ErtsMigrationPath *mp = &new_mpaths->mpath[qix];
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+	mp->sched_util = sched_util_balancing;
+#endif
 	mp->flags = flags;
 	mp->misc_evac_runq = NULL;
 
@@ -4628,6 +4955,11 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online)
     set_wakeup_other_data();
 #endif
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    if (erts_sched_balance_util)
+	erts_sched_compact_load = 0;
+#endif
+
     ASSERT(no_schedulers_online <= no_schedulers);
     ASSERT(no_schedulers_online >= 1);
     ASSERT(no_schedulers >= 1);
@@ -4696,6 +5028,11 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online)
 	rq->ports.info.reds = 0;
 	rq->ports.start = NULL;
 	rq->ports.end = NULL;
+
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+	init_runq_sched_util(&rq->sched_util, erts_sched_balance_util);
+#endif
+
     }
 
 #ifdef ERTS_SMP
@@ -4794,6 +5131,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online)
 	esdp->reductions = 0;
 
 	init_sched_wall_time(&esdp->sched_wall_time);
+
 	erts_port_task_handle_init(&esdp->nosuspend_port_task_handle);
 
     }
@@ -5761,7 +6099,7 @@ erts_set_schedulers_online(Process *p,
 
 		    for (ix = no; ix < online; ix++) {
 			ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
-			wake_scheduler(rq, 0);
+			wake_scheduler(rq);
 		    }
 		}
 	    }
@@ -5860,7 +6198,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all)
 
 		for (ix = 1; ix < online; ix++) {
 		    ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
-		    wake_scheduler(rq, 0);
+		    wake_scheduler(rq);
 		}
 
 		if (erts_smp_atomic32_read_nob(&schdlr_sspnd.active)
@@ -7294,7 +7632,7 @@ Process *schedule(Process *p, int calls)
     continue_check_activities_to_run:
 	flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
     continue_check_activities_to_run_known_flags:
-
+	ASSERT(flags & ERTS_RUNQ_FLG_NONEMPTY);
 
 	if (flags & (ERTS_RUNQ_FLG_CHK_CPU_BIND|ERTS_RUNQ_FLG_SUSPENDED)) {
 	
@@ -7346,20 +7684,16 @@ Process *schedule(Process *p, int calls)
 	    rq->wakeup_other = 0;
 	    rq->wakeup_other_reds = 0;
 
-	    empty_runq(rq);
-
 	    flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
-	    if (flags & ERTS_RUNQ_FLG_SUSPENDED) {
-		non_empty_runq(rq);
+	    if (flags & ERTS_RUNQ_FLG_SUSPENDED)
 		goto continue_check_activities_to_run_known_flags;
-	    }
-	    else if (!(flags & ERTS_RUNQ_FLG_INACTIVE)) {
-		if (try_steal_task(rq)) {
-		    non_empty_runq(rq);
+	    if (flags & ERTS_RUNQ_FLG_INACTIVE)
+		empty_runq(rq);
+	    else {
+		if (try_steal_task(rq))
 		    goto continue_check_activities_to_run;
-		}
 
-		(void) ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_PROTECTED);
+		empty_runq(rq);
 
 		/*
 		 * Check for ERTS_RUNQ_FLG_SUSPENDED has to be done
@@ -7371,7 +7705,6 @@ Process *schedule(Process *p, int calls)
 		    goto continue_check_activities_to_run_known_flags;
 		}
 	    }
-
 #endif
 
 	    scheduler_wait(&fcalls, esdp, rq);
@@ -8486,6 +8819,10 @@ erts_schedule_misc_op(void (*func)(void *), void *arg)
 	rq->misc.start = molp;
     rq->misc.end = molp;
 
+#ifdef ERTS_SMP
+    non_empty_runq(rq);
+#endif
+
     erts_smp_runq_unlock(rq);
 
     smp_notify_inc_runq(rq);
@@ -9372,8 +9709,11 @@ save_pending_exiter(Process *p)
 
     erts_proclist_store_last(&rq->procs.pending_exiters, plp);
 
+    non_empty_runq(rq);
+
     erts_smp_runq_unlock(rq);
-    wake_scheduler(rq, 1);
+
+    wake_scheduler(rq);
 }
 
 #endif
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 043621125c..c2ef55f2d5 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -70,6 +70,9 @@ typedef struct process Process;
 
 struct ErtsNodesMonitor_;
 
+#define ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT_OPT	0
+#define ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT		0
+
 #define ERTS_MAX_NO_OF_SCHEDULERS 1024
 
 #define ERTS_DEFAULT_MAX_PROCESSES (1 << 18)
@@ -98,6 +101,7 @@ struct saved_calls {
 
 extern Export exp_send, exp_receive, exp_timeout;
 extern int erts_sched_compact_load;
+extern int erts_sched_balance_util;
 extern Uint erts_no_schedulers;
 extern Uint erts_no_run_queues;
 extern int erts_sched_thread_suggested_stack_size;
@@ -198,6 +202,10 @@ extern int erts_sched_thread_suggested_stack_size;
 #define ERTS_RUNQ_FLGS_SET(RQ, FLGS)					\
     ((Uint32) erts_smp_atomic32_read_bor_relb(&(RQ)->flags,		\
 					      (erts_aint32_t) (FLGS)))
+#define ERTS_RUNQ_FLGS_BSET(RQ, MSK, FLGS)				\
+    ((Uint32) erts_smp_atomic32_read_bset_relb(&(RQ)->flags,		\
+					       (erts_aint32_t) (MSK),	\
+					       (erts_aint32_t) (FLGS)))
 #define ERTS_RUNQ_FLGS_UNSET(RQ, FLGS)					\
     ((Uint32) erts_smp_atomic32_read_band_relb(&(RQ)->flags,		\
 					       (erts_aint32_t) ~(FLGS)))
@@ -316,9 +324,40 @@ typedef struct {
     int reds;
 } ErtsRunQueueInfo;
 
+
+#ifdef HAVE_GETHRTIME
+#  undef ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT_OPT
+#  define ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT_OPT 1
+#endif
+
 #ifdef ERTS_SMP
 
+#undef ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+#define ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT_OPT
+
+#ifdef ARCH_64
+typedef erts_atomic_t ErtsAtomicSchedTime;
+#elif defined(ARCH_32)
+typedef erts_dw_atomic_t ErtsAtomicSchedTime;
+#else
+# error :-/
+#endif
+
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+typedef struct {
+    ErtsAtomicSchedTime last;
+    struct {
+	Uint64 short_interval;
+	Uint64 long_interval;
+    } worktime;
+    int is_working;
+} ErtsRunQueueSchedUtil;
+#endif
+
 typedef struct {
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    int sched_util;
+#endif
     Uint32 flags;
     ErtsRunQueue *misc_evac_runq;
     struct {
@@ -385,6 +424,9 @@ struct ErtsRunQueue_ {
 	Port *start;
 	Port *end;
     } ports;
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+    ErtsRunQueueSchedUtil sched_util;
+#endif
 };
 
 #ifdef ERTS_SMP
@@ -414,6 +456,7 @@ do {								\
 } while (0)
 
 typedef struct {
+    int need; /* "+sbu true" or scheduler_wall_time enabled */
     int enabled;
     Uint64 start;
     struct {
@@ -542,6 +585,12 @@ int erts_smp_lc_runq_is_locked(ErtsRunQueue *);
 
 #ifdef ERTS_INCLUDE_SCHEDULER_INTERNALS
 
+#ifdef ERTS_SMP
+void erts_empty_runq(ErtsRunQueue *rq);
+void erts_non_empty_runq(ErtsRunQueue *rq);
+#endif
+
+
 /*
  * Run queue locked during modifications. We use atomic ops since
  * other threads peek at values without run queue lock.
@@ -574,6 +623,10 @@ erts_smp_inc_runq_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi, int prio)
 
     erts_smp_atomic32_set_relb(&rqi->len, len);
 
+#ifdef ERTS_SMP
+    if (rq->len == 0)
+	erts_non_empty_runq(rq);
+#endif
     rq->len++;
     if (rq->max_len < rq->len)
 	rq->max_len = len;
@@ -1695,6 +1748,13 @@ erts_proc_set_error_handler(Process *p, ErtsProcLocks plocks, Eterm handler)
 
 extern erts_atomic_t erts_migration_paths;
 
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+int erts_get_sched_util(ErtsRunQueue *rq,
+			int initially_locked,
+			int short_interval);
+#endif
+
+
 ERTS_GLB_INLINE ErtsMigrationPaths *erts_get_migration_paths_managed(void);
 ERTS_GLB_INLINE ErtsMigrationPaths *erts_get_migration_paths(void);
 ERTS_GLB_INLINE ErtsRunQueue *erts_check_emigration_need(ErtsRunQueue *c_rq,
@@ -1746,22 +1806,36 @@ erts_check_emigration_need(ErtsRunQueue *c_rq, int prio)
 		return mp->prio[prio].runq;
 	}
 
-
-	if (prio == ERTS_PORT_PRIO_LEVEL)
-	    len = RUNQ_READ_LEN(&c_rq->ports.info.len);
+#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
+	if (mp->sched_util) {
+	    ErtsRunQueue *rq = mp->prio[prio].runq;
+	    /* No migration if other is non-empty */
+	    if (!(ERTS_RUNQ_FLGS_GET(rq) & ERTS_RUNQ_FLG_NONEMPTY)
+		&& erts_get_sched_util(rq, 0, 1) < mp->prio[prio].limit.other
+		&& erts_get_sched_util(c_rq, 0, 1) > mp->prio[prio].limit.this) {
+		return rq;
+	    }
+	}
 	else
-	    len = RUNQ_READ_LEN(&c_rq->procs.prio_info[prio].len);
-
-	if (len > mp->prio[prio].limit.this) {
-	    ErtsRunQueue *n_rq = mp->prio[prio].runq;
-	    if (n_rq) {
-		if (prio == ERTS_PORT_PRIO_LEVEL)
-		    len = RUNQ_READ_LEN(&n_rq->ports.info.len);
-		else
-		    len = RUNQ_READ_LEN(&n_rq->procs.prio_info[prio].len);
-
-		if (len < mp->prio[prio].limit.other)
-		    return n_rq;
+#endif
+	{
+
+	    if (prio == ERTS_PORT_PRIO_LEVEL)
+		len = RUNQ_READ_LEN(&c_rq->ports.info.len);
+	    else
+		len = RUNQ_READ_LEN(&c_rq->procs.prio_info[prio].len);
+
+	    if (len > mp->prio[prio].limit.this) {
+		ErtsRunQueue *n_rq = mp->prio[prio].runq;
+		if (n_rq) {
+		    if (prio == ERTS_PORT_PRIO_LEVEL)
+			len = RUNQ_READ_LEN(&n_rq->ports.info.len);
+		    else
+			len = RUNQ_READ_LEN(&n_rq->procs.prio_info[prio].len);
+
+		    if (len < mp->prio[prio].limit.other)
+			return n_rq;
+		}
 	    }
 	}
     }