X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=steadystate.c;h=43c715c96422dddfcbca6f8fed9871c80ee4ffd2;hp=3b9b6ef7461835f78a6d82b2c45fbadc97b0b6c5;hb=4c14c109614a76bdbde1d030f7129324f1472a1f;hpb=6da94b076ca08f884dc83c9978bc995848dea0e8

diff --git a/steadystate.c b/steadystate.c
index 3b9b6ef7..43c715c9 100644
--- a/steadystate.c
+++ b/steadystate.c
@@ -4,12 +4,22 @@
 #include "steadystate.h"
 #include "helper_thread.h"
 
-void steadystate_setup()
+bool steadystate_enabled = false;
+
+static void steadystate_alloc(struct thread_data *td)
+{
+	td->ss.bw_data = calloc(td->ss.dur, sizeof(uint64_t));
+	td->ss.iops_data = calloc(td->ss.dur, sizeof(uint64_t));
+
+	td->ss.state |= __FIO_SS_DATA;
+}
+
+void steadystate_setup(void)
 {
 	int i, prev_groupid;
 	struct thread_data *td, *prev_td;
 
-	if (!steadystate)
+	if (!steadystate_enabled)
 		return;
 
 	/*
@@ -20,7 +30,7 @@ void steadystate_setup()
 	prev_groupid = -1;
 	prev_td = NULL;
 	for_each_td(td, i) {
-		if (td->ts.ss == NULL)
+		if (!td->ss.dur)
 			continue;
 
 		if (!td->o.group_reporting) {
@@ -30,7 +40,6 @@ void steadystate_setup()
 
 		if (prev_groupid != td->groupid) {
 			if (prev_td != NULL) {
-				prev_td->ss.last_in_group = 1;
 				steadystate_alloc(prev_td);
 			}
 			prev_groupid = td->groupid;
@@ -39,37 +48,166 @@ void steadystate_setup()
 	}
 
 	if (prev_td != NULL && prev_td->o.group_reporting) {
-		prev_td->ss.last_in_group = 1;
 		steadystate_alloc(prev_td);
 	}
 }
 
-void steadystate_alloc(struct thread_data *td)
+static bool steadystate_slope(uint64_t iops, uint64_t bw,
+			      struct thread_data *td)
+{
+	int i, j;
+	double result;
+	struct steadystate_data *ss = &td->ss;
+	uint64_t new_val;
+
+	ss->bw_data[ss->tail] = bw;
+	ss->iops_data[ss->tail] = iops;
+
+	if (ss->state & __FIO_SS_IOPS)
+		new_val = iops;
+	else
+		new_val = bw;
+
+	if (ss->state & __FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
+		if (!(ss->state & __FIO_SS_BUFFER_FULL)) {
+			/* first time through */
+			for(i = 0, ss->sum_y = 0; i < ss->dur; i++) {
+				if (ss->state & __FIO_SS_IOPS)
+					ss->sum_y += ss->iops_data[i];
+				else
+					ss->sum_y += ss->bw_data[i];
+				j = (ss->head + i) % ss->dur;
+				if (ss->state & __FIO_SS_IOPS)
+					ss->sum_xy += i * ss->iops_data[j];
+				else
+					ss->sum_xy += i * ss->bw_data[j];
+			}
+			ss->state |= __FIO_SS_BUFFER_FULL;
+		} else {		/* easy to update the sums */
+			ss->sum_y -= ss->oldest_y;
+			ss->sum_y += new_val;
+			ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
+		}
+
+		if (ss->state & __FIO_SS_IOPS)
+			ss->oldest_y = ss->iops_data[ss->head];
+		else
+			ss->oldest_y = ss->bw_data[ss->head];
+
+		/*
+		 * calculate slope as (sum_xy - sum_x * sum_y / n) / (sum_(x^2)
+		 * - (sum_x)^2 / n) This code assumes that all x values are
+		 * equally spaced when they are often off by a few milliseconds.
+		 * This assumption greatly simplifies the calculations.
+		 */
+		ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) /
+				(ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
+		if (ss->state & __FIO_SS_PCT)
+			ss->criterion = 100.0 * ss->slope / (ss->sum_y / ss->dur);
+		else
+			ss->criterion = ss->slope;
+
+		dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
+					"criterion: %f, limit: %f\n",
+					(unsigned long long) ss->sum_y,
+					(unsigned long long) ss->sum_xy,
+					ss->slope, ss->criterion, ss->limit);
+
+		result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
+		if (result < ss->limit)
+			return true;
+	}
+
+	ss->tail = (ss->tail + 1) % ss->dur;
+	if (ss->tail <= ss->head)
+		ss->head = (ss->head + 1) % ss->dur;
+
+	return false;
+}
+
+static bool steadystate_deviation(uint64_t iops, uint64_t bw,
+				  struct thread_data *td)
 {
 	int i;
+	double diff;
+	double mean;
+
+	struct steadystate_data *ss = &td->ss;
+
+	ss->bw_data[ss->tail] = bw;
+	ss->iops_data[ss->tail] = iops;
+
+	if (ss->state & __FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
+		if (!(ss->state & __FIO_SS_BUFFER_FULL)) {
+			/* first time through */
+			for(i = 0, ss->sum_y = 0; i < ss->dur; i++)
+				if (ss->state & __FIO_SS_IOPS)
+					ss->sum_y += ss->iops_data[i];
+				else
+					ss->sum_y += ss->bw_data[i];
+			ss->state |= __FIO_SS_BUFFER_FULL;
+		} else {		/* easy to update the sum */
+			ss->sum_y -= ss->oldest_y;
+			if (ss->state & __FIO_SS_IOPS)
+				ss->sum_y += ss->iops_data[ss->tail];
+			else
+				ss->sum_y += ss->bw_data[ss->tail];
+		}
+
+		if (ss->state & __FIO_SS_IOPS)
+			ss->oldest_y = ss->iops_data[ss->head];
+		else
+			ss->oldest_y = ss->bw_data[ss->head];
+
+		mean = (double) ss->sum_y / ss->dur;
+		ss->deviation = 0.0;
+
+		for (i = 0; i < ss->dur; i++) {
+			if (ss->state & __FIO_SS_IOPS)
+				diff = ss->iops_data[i] - mean;
+			else
+				diff = ss->bw_data[i] - mean;
+			ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
+		}
+
+		if (ss->state & __FIO_SS_PCT)
+			ss->criterion = 100.0 * ss->deviation / mean;
+		else
+			ss->criterion = ss->deviation;
+
+		dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
+					"objective: %f, limit: %f\n",
+					(unsigned long long) ss->sum_y, mean,
+					ss->deviation, ss->criterion, ss->limit);
+
+		if (ss->criterion < ss->limit)
+			return true;
+	}
+
+	ss->tail = (ss->tail + 1) % ss->dur;
+	if (ss->tail <= ss->head)
+		ss->head = (ss->head + 1) % ss->dur;
 
-	td->ss.bw_data = malloc(td->ss.dur * sizeof(unsigned long));
-	td->ss.iops_data = malloc(td->ss.dur * sizeof(unsigned long));
-	/* initialize so that it is obvious if the cache is not full in the output */
-	for (i = 0; i < td->ss.dur; i++)
-		td->ss.iops_data[i] = td->ss.bw_data[i] = 0;
+	return false;
 }
 
-void steadystate_check()
+void steadystate_check(void)
 {
 	int i, j, ddir, prev_groupid, group_ramp_time_over = 0;
 	unsigned long rate_time;
 	struct thread_data *td, *td2;
 	struct timeval now;
-	unsigned long group_bw = 0, group_iops = 0;
-	unsigned long long td_iops;
-	unsigned long long td_bytes;
+	uint64_t group_bw = 0, group_iops = 0;
+	uint64_t td_iops, td_bytes;
+	bool ret;
 
 	prev_groupid = -1;
 	for_each_td(td, i) {
 		struct steadystate_data *ss = &td->ss;
 
-		if (!ss->dur || td->runstate <= TD_SETTING_UP || td->runstate >= TD_EXITED || ss->attained)
+		if (!ss->dur || td->runstate <= TD_SETTING_UP ||
+		    td->runstate >= TD_EXITED || !ss->state ||
+		    ss->state & __FIO_SS_ATTAINED)
 			continue;
 
 		td_iops = 0;
@@ -83,13 +221,14 @@ void steadystate_check()
 		prev_groupid = td->groupid;
 
 		fio_gettime(&now, NULL);
-		if (ss->ramp_time && !ss->ramp_time_over)
-			/* 
+		if (ss->ramp_time && !(ss->state & __FIO_SS_RAMP_OVER)) {
+			/*
 			 * Begin recording data one second after ss->ramp_time
 			 * has elapsed
 			 */
 			if (utime_since(&td->epoch, &now) >= (ss->ramp_time + 1000000L))
-				ss->ramp_time_over = 1;
+				ss->state |= __FIO_SS_RAMP_OVER;
+		}
 
 		td_io_u_lock(td);
 		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
@@ -101,143 +240,124 @@ void steadystate_check()
 		rate_time = mtime_since(&ss->prev_time, &now);
 		memcpy(&ss->prev_time, &now, sizeof(now));
 
-		/* 
+		/*
 		 * Begin monitoring when job starts but don't actually use
 		 * data in checking stopping criterion until ss->ramp_time is
 		 * over. This ensures that we will have a sane value in
 		 * prev_iops/bw the first time through after ss->ramp_time
 		 * is done.
 		 */
-		if (ss->ramp_time_over) {
+		if (ss->state & __FIO_SS_RAMP_OVER) {
 			group_bw += 1000 * (td_bytes - ss->prev_bytes) / rate_time;
-	        	group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
+			group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
 			++group_ramp_time_over;
 		}
 		ss->prev_iops = td_iops;
 		ss->prev_bytes = td_bytes;
 
-		if (td->o.group_reporting && !ss->last_in_group)
+		if (td->o.group_reporting && !(ss->state & __FIO_SS_DATA))
 			continue;
 
-		/* don't begin checking criterion until ss->ramp_time is over for at least one thread in group */
+		/*
+		 * Don't begin checking criterion until ss->ramp_time is over
+		 * for at least one thread in group
+		 */
 		if (!group_ramp_time_over)
 			continue;
 
-		dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, groupid: %u, rate_msec: %ld, iops: %lu, bw: %lu, head: %d, tail: %d\n", 
-			i, td->groupid, rate_time, group_iops, group_bw, ss->head, ss->tail);
-
-		if (ss->evaluate(group_iops, group_bw, td))
-		{
-			if (td->o.group_reporting)
+		dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
+					"groupid: %u, rate_msec: %ld, "
+					"iops: %llu, bw: %llu, head: %d, tail: %d\n",
+					i, td->groupid, rate_time,
+					(unsigned long long) group_iops,
+					(unsigned long long) group_bw,
+					ss->head, ss->tail);
+
+		if (ss->state & __FIO_SS_SLOPE)
+			ret = steadystate_slope(group_iops, group_bw, td);
+		else
+			ret = steadystate_deviation(group_iops, group_bw, td);
+
+		if (ret) {
+			if (td->o.group_reporting) {
 				for_each_td(td2, j) {
 					if (td2->groupid == td->groupid) {
-						td2->ss.attained = 1;
+						td2->ss.state |= __FIO_SS_ATTAINED;
 						fio_mark_td_terminate(td2);
 					}
 				}
-			else {
-				ss->attained = 1;
+			} else {
+				ss->state |= __FIO_SS_ATTAINED;
 				fio_mark_td_terminate(td);
 			}
 		}
 	}
 }
 
-bool steadystate_slope(unsigned long iops, unsigned long bw, struct thread_data *td)
+int td_steadystate_init(struct thread_data *td)
 {
-	int i, j;
-	double result;
 	struct steadystate_data *ss = &td->ss;
-	unsigned long new_val = ss->check_iops ? iops : bw;
+	struct thread_options *o = &td->o;
+	struct thread_data *td2;
+	int j;
 
-	ss->bw_data[ss->tail] = bw;
-	ss->iops_data[ss->tail] = iops;
+	memset(ss, 0, sizeof(*ss));
 
-	if (ss->tail < ss->head || (ss->tail - ss->head == ss->dur - 1))
-	{
-		if (ss->sum_y == 0)	/* first time through */
-		{
-			for(i = 0; i < ss->dur; i++)
-			{
-				ss->sum_y += ss->check_iops ? ss->iops_data[i] : ss->bw_data[i];
-				j = ss->head + i;
-				if (j >= ss->dur)
-					j -= ss->dur;
-				ss->sum_xy += (ss->check_iops ? ss->iops_data[j] : ss->bw_data[j]) * i;
-			}
-		} else {		/* easy to update the sums */
-			ss->sum_y -= ss->oldest_y;
-			ss->sum_y += new_val;
-			ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
-		}
+	if (o->ss_dur) {
+		steadystate_enabled = true;
+		o->ss_dur /= 1000000L;
 
-		ss->oldest_y = ss->check_iops ? ss->iops_data[ss->head] : ss->bw_data[ss->head];
+		/* put all steady state info in one place */
+		ss->dur = o->ss_dur;
+		ss->limit = o->ss_limit.u.f;
+		ss->ramp_time = o->ss_ramp_time;
 
-		/*
-		 * calculate slope as (sum_xy - sum_x * sum_y / n) / (sum_(x^2) - (sum_x)^2 / n)
-		 * This code assumes that all x values are equally spaced when they are often
-		 * off by a few milliseconds. This assumption greatly simplifies the
-		 * calculations.
-		 */
-		ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) / (ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
-		ss->criterion = ss->pct ? 100.0 * ss->slope / (ss->sum_y / ss->dur) : ss->slope;
+		ss->state = o->ss_state;
+		if (!td->ss.ramp_time)
+			ss->state |= __FIO_SS_RAMP_OVER;
 
-		dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, criterion: %f, limit: %f\n",
-			ss->sum_y, ss->sum_xy, ss->slope, ss->criterion, ss->limit);
+		ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
+		ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
+	}
 
-		result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
-		if (result < ss->limit)
-			return true;
+	/* make sure that ss options are consistent within reporting group */
+	for_each_td(td2, j) {
+		if (td2->groupid == td->groupid) {
+			struct steadystate_data *ss2 = &td2->ss;
+
+			if (ss2->dur != ss->dur ||
+			    ss2->limit != ss->limit ||
+			    ss2->ramp_time != ss->ramp_time ||
+			    ss2->state != ss->state ||
+			    ss2->sum_x != ss->sum_x ||
+			    ss2->sum_x_sq != ss->sum_x_sq) {
+				td_verror(td, EINVAL, "job rejected: steadystate options must be consistent within reporting groups");
+				return 1;
+			}
+		}
 	}
 
-	ss->tail = (ss->tail + 1) % ss->dur;
-	if (ss->tail <= ss->head)
-		ss->head = (ss->head + 1) % ss->dur;
-	return false;
+	return 0;
 }
 
-bool steadystate_deviation(unsigned long iops, unsigned long bw, struct thread_data *td)
+uint64_t steadystate_bw_mean(struct thread_stat *ts)
 {
 	int i;
-	double diff;
-	double mean;
+	uint64_t sum;
 
-	struct steadystate_data *ss = &td->ss;
+	for (i = 0, sum = 0; i < ts->ss_dur; i++)
+		sum += ts->ss_bw_data[i];
 
-	ss->bw_data[ss->tail] = bw;
-	ss->iops_data[ss->tail] = iops;
-
-	if (ss->tail < ss->head || (ss->tail - ss->head == ss->dur - 1))
-	{
-		if (ss->sum_y == 0)	/* first time through */
-		{
-			for(i = 0; i < ss->dur; i++)
-				ss->sum_y += ss->check_iops ? ss->iops_data[i] : ss->bw_data[i];
-		} else {		/* easy to update the sum */
-			ss->sum_y -= ss->oldest_y;
-			ss->sum_y += ss->check_iops ? ss->iops_data[ss->tail] : ss->bw_data[ss->tail];
-		}
-
-		ss->oldest_y = ss->check_iops ? ss->iops_data[ss->head] : ss->bw_data[ss->head];
-		mean = (double) ss->sum_y / ss->dur;
-		ss->deviation = 0.0;
-
-		for (i = 0; i < ss->dur; i++)
-		{	
-			diff = (double) (ss->check_iops ? ss->iops_data[i] : ss->bw_data[i]) - mean;
-			ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
-		}
-
-		ss->criterion = ss->pct ? 100.0 * ss->deviation / mean : ss->deviation;
+	return sum / ts->ss_dur;
+}
 
-		dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, objective: %f, limit: %f\n", ss->sum_y, mean, ss->deviation, ss->criterion, ss->limit);
+uint64_t steadystate_iops_mean(struct thread_stat *ts)
+{
+	int i;
+	uint64_t sum;
 
-		if (ss->criterion < ss->limit)
-			return true;
-	}
+	for (i = 0, sum = 0; i < ts->ss_dur; i++)
+		sum += ts->ss_iops_data[i];
 
-	ss->tail = (ss->tail + 1) % ss->dur;
-	if (ss->tail <= ss->head)
-		ss->head = (ss->head + 1) % ss->dur;
-	return false;
+	return sum / ts->ss_dur;
 }