[fio.git] / steadystate.c

#include <stdlib.h>

#include "fio.h"
#include "steadystate.h"
#include "helper_thread.h"

bool steadystate_enabled = false;

void steadystate_setup(void)
{
	int i, prev_groupid;
	struct thread_data *td, *prev_td;

	if (!steadystate_enabled)
		return;

	/*
	 * if group reporting is enabled, identify the last td
	 * for each group and use it for storing steady state
	 * data
	 */
	prev_groupid = -1;
	prev_td = NULL;
	for_each_td(td, i) {
		if (td->ts.ss == NULL)
			continue;

		if (!td->o.group_reporting) {
			steadystate_alloc(td);
			continue;
		}

		if (prev_groupid != td->groupid) {
			if (prev_td != NULL) {
				prev_td->ss.last_in_group = 1;
				steadystate_alloc(prev_td);
			}
			prev_groupid = td->groupid;
		}
		prev_td = td;
	}

	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)
{
	int i;

	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;
}

static bool steadystate_slope(unsigned long iops, unsigned long bw,
			      struct thread_data *td)
{
	int i, j;
	double result;
	struct steadystate_data *ss = &td->ss;
	unsigned long new_val;

	ss->bw_data[ss->tail] = bw;
	ss->iops_data[ss->tail] = iops;

	if (ss->check_iops)
		new_val = iops;
	else
		new_val = bw;

	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++) {
				if (ss->check_iops)
					ss->sum_y += ss->iops_data[i];
				else
					ss->sum_y += ss->bw_data[i];
				j = ss->head + i;
				if (j >= ss->dur)
					j -= ss->dur;
				if (ss->check_iops)
					ss->sum_xy += ss->iops_data[j];
				else
					ss->sum_xy += ss->bw_data[j];
			}
		} 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->check_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->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",
					ss->sum_y, 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(unsigned long iops, unsigned long 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->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++)
				if (ss->check_iops)
					ss->sum_y += ss->iops_data[i];
				else
					ss->sum_y += ss->bw_data[i];
		} else {		/* easy to update the sum */
			ss->sum_y -= ss->oldest_y;
			if (ss->check_iops)
				ss->sum_y += ss->iops_data[ss->tail];
			else
				ss->sum_y += ss->bw_data[ss->tail];
		}

		if (ss->check_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->check_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->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",
					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;

	return false;
}

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;
	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)
			continue;

		td_iops = 0;
		td_bytes = 0;
		if (!td->o.group_reporting ||
		    (td->o.group_reporting && td->groupid != prev_groupid)) {
			group_bw = 0;
			group_iops = 0;
			group_ramp_time_over = 0;
		}
		prev_groupid = td->groupid;

		fio_gettime(&now, NULL);
		if (ss->ramp_time && !ss->ramp_time_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;

		td_io_u_lock(td);
		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
			td_iops += td->io_blocks[ddir];
			td_bytes += td->io_bytes[ddir];
		}
		td_io_u_unlock(td);

		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) {
			group_bw += 1000 * (td_bytes - ss->prev_bytes) / 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)
			continue;

		/*
		 * 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 (steadystate_check_slope(&td->o))
			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;
						fio_mark_td_terminate(td2);
					}
				}
			} else {
				ss->attained = 1;
				fio_mark_td_terminate(td);
			}
		}
	}
}
Commit	Line	Data
ba8fb6f6 VF	1	#include <stdlib.h>
ba8fb6f6 VF	2
16e56d25 VF	3	#include "fio.h"
	4	#include "steadystate.h"
	5	#include "helper_thread.h"
	6
84784e07	7	bool steadystate_enabled = false;
4cd5b75b JA	8
4cd5b75b JA	9	void steadystate_setup(void)
16e56d25 VF	10	{
	11	int i, prev_groupid;
	12	struct thread_data td, prev_td;
	13
84784e07	14	if (!steadystate_enabled)
16e56d25 VF	15	return;
	16
	17	/*
	18	* if group reporting is enabled, identify the last td
	19	* for each group and use it for storing steady state
	20	* data
	21	*/
	22	prev_groupid = -1;
	23	prev_td = NULL;
	24	for_each_td(td, i) {
ba8fb6f6 VF	25	if (td->ts.ss == NULL)
	26	continue;
	27
	28	if (!td->o.group_reporting) {
	29	steadystate_alloc(td);
16e56d25	30	continue;
ba8fb6f6	31	}
16e56d25 VF	32
16e56d25 VF	33	if (prev_groupid != td->groupid) {
ba8fb6f6	34	if (prev_td != NULL) {
16e56d25	35	prev_td->ss.last_in_group = 1;
ba8fb6f6 VF	36	steadystate_alloc(prev_td);
ba8fb6f6 VF	37	}
16e56d25 VF	38	prev_groupid = td->groupid;
	39	}
	40	prev_td = td;
	41	}
	42
ba8fb6f6	43	if (prev_td != NULL && prev_td->o.group_reporting) {
16e56d25	44	prev_td->ss.last_in_group = 1;
ba8fb6f6 VF	45	steadystate_alloc(prev_td);
	46	}
	47	}
	48
	49	void steadystate_alloc(struct thread_data *td)
	50	{
	51	int i;
	52
	53	td->ss.bw_data = malloc(td->ss.dur * sizeof(unsigned long));
	54	td->ss.iops_data = malloc(td->ss.dur * sizeof(unsigned long));
	55	/* initialize so that it is obvious if the cache is not full in the output */
	56	for (i = 0; i < td->ss.dur; i++)
	57	td->ss.iops_data[i] = td->ss.bw_data[i] = 0;
16e56d25 VF	58	}
16e56d25 VF	59
4cd5b75b JA	60	static bool steadystate_slope(unsigned long iops, unsigned long bw,
4cd5b75b JA	61	struct thread_data *td)
16e56d25	62	{
ba8fb6f6	63	int i, j;
16e56d25	64	double result;
16e56d25	65	struct steadystate_data *ss = &td->ss;
7be78a8a	66	unsigned long new_val;
16e56d25	67
ba8fb6f6 VF	68	ss->bw_data[ss->tail] = bw;
ba8fb6f6 VF	69	ss->iops_data[ss->tail] = iops;
16e56d25	70
7be78a8a JA	71	if (ss->check_iops)
	72	new_val = iops;
	73	else
	74	new_val = bw;
	75
4cd5b75b JA	76	if (ss->tail < ss->head \|\| (ss->tail - ss->head == ss->dur - 1)) {
	77	if (ss->sum_y == 0) { /* first time through */
	78	for(i = 0; i < ss->dur; i++) {
7be78a8a JA	79	if (ss->check_iops)
	80	ss->sum_y += ss->iops_data[i];
	81	else
	82	ss->sum_y += ss->bw_data[i];
ba8fb6f6 VF	83	j = ss->head + i;
	84	if (j >= ss->dur)
	85	j -= ss->dur;
7be78a8a JA	86	if (ss->check_iops)
	87	ss->sum_xy += ss->iops_data[j];
	88	else
	89	ss->sum_xy += ss->bw_data[j];
16e56d25 VF	90	}
	91	} else { /* easy to update the sums */
	92	ss->sum_y -= ss->oldest_y;
ba8fb6f6 VF	93	ss->sum_y += new_val;
ba8fb6f6 VF	94	ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
16e56d25 VF	95	}
16e56d25 VF	96
7be78a8a JA	97	if (ss->check_iops)
	98	ss->oldest_y = ss->iops_data[ss->head];
	99	else
	100	ss->oldest_y = ss->bw_data[ss->head];
16e56d25 VF	101
16e56d25 VF	102	/*
7be78a8a JA	103	* calculate slope as (sum_xy - sum_x * sum_y / n) / (sum_(x^2)
	104	* - (sum_x)^2 / n) This code assumes that all x values are
	105	* equally spaced when they are often off by a few milliseconds.
	106	* This assumption greatly simplifies the calculations.
16e56d25	107	*/
7be78a8a JA	108	ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) /
	109	(ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
	110	if (ss->pct)
	111	ss->criterion = 100.0 * ss->slope / (ss->sum_y / ss->dur);
	112	else
	113	ss->criterion = ss->slope;
16e56d25	114
7be78a8a JA	115	dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
	116	"criterion: %f, limit: %f\n",
	117	ss->sum_y, ss->sum_xy, ss->slope,
	118	ss->criterion, ss->limit);
16e56d25	119
6da94b07	120	result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
16e56d25 VF	121	if (result < ss->limit)
	122	return true;
	123	}
	124
	125	ss->tail = (ss->tail + 1) % ss->dur;
	126	if (ss->tail <= ss->head)
	127	ss->head = (ss->head + 1) % ss->dur;
4cd5b75b	128
16e56d25 VF	129	return false;
	130	}
	131
4cd5b75b JA	132	static bool steadystate_deviation(unsigned long iops, unsigned long bw,
4cd5b75b JA	133	struct thread_data *td)
16e56d25 VF	134	{
	135	int i;
	136	double diff;
	137	double mean;
16e56d25 VF	138
	139	struct steadystate_data *ss = &td->ss;
	140
ba8fb6f6 VF	141	ss->bw_data[ss->tail] = bw;
ba8fb6f6 VF	142	ss->iops_data[ss->tail] = iops;
16e56d25	143
4cd5b75b JA	144	if (ss->tail < ss->head \|\| (ss->tail - ss->head == ss->dur - 1)) {
4cd5b75b JA	145	if (ss->sum_y == 0) { /* first time through */
16e56d25	146	for(i = 0; i < ss->dur; i++)
7be78a8a JA	147	if (ss->check_iops)
	148	ss->sum_y += ss->iops_data[i];
	149	else
	150	ss->sum_y += ss->bw_data[i];
16e56d25 VF	151	} else { /* easy to update the sum */
16e56d25 VF	152	ss->sum_y -= ss->oldest_y;
7be78a8a JA	153	if (ss->check_iops)
	154	ss->sum_y += ss->iops_data[ss->tail];
	155	else
	156	ss->sum_y += ss->bw_data[ss->tail];
16e56d25 VF	157	}
16e56d25 VF	158
7be78a8a JA	159	if (ss->check_iops)
	160	ss->oldest_y = ss->iops_data[ss->head];
	161	else
	162	ss->oldest_y = ss->bw_data[ss->head];
	163
16e56d25	164	mean = (double) ss->sum_y / ss->dur;
ba8fb6f6	165	ss->deviation = 0.0;
16e56d25	166
4cd5b75b	167	for (i = 0; i < ss->dur; i++) {
7be78a8a JA	168	if (ss->check_iops)
	169	diff = ss->iops_data[i] - mean;
	170	else
	171	diff = ss->bw_data[i] - mean;
6da94b07	172	ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
16e56d25 VF	173	}
16e56d25 VF	174
7be78a8a JA	175	if (ss->pct)
	176	ss->criterion = 100.0 * ss->deviation / mean;
	177	else
	178	ss->criterion = ss->deviation;
16e56d25	179
7be78a8a JA	180	dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
	181	"objective: %f, limit: %f\n",
	182	ss->sum_y, mean, ss->deviation,
	183	ss->criterion, ss->limit);
16e56d25 VF	184
	185	if (ss->criterion < ss->limit)
	186	return true;
	187	}
	188
	189	ss->tail = (ss->tail + 1) % ss->dur;
	190	if (ss->tail <= ss->head)
	191	ss->head = (ss->head + 1) % ss->dur;
4cd5b75b	192
16e56d25 VF	193	return false;
16e56d25 VF	194	}
4cd5b75b JA	195
	196	void steadystate_check(void)
	197	{
	198	int i, j, ddir, prev_groupid, group_ramp_time_over = 0;
	199	unsigned long rate_time;
	200	struct thread_data td, td2;
	201	struct timeval now;
	202	unsigned long group_bw = 0, group_iops = 0;
	203	unsigned long long td_iops;
	204	unsigned long long td_bytes;
	205	bool ret;
	206
	207	prev_groupid = -1;
	208	for_each_td(td, i) {
	209	struct steadystate_data *ss = &td->ss;
	210
7be78a8a JA	211	if (!ss->dur \|\| td->runstate <= TD_SETTING_UP \|\|
7be78a8a JA	212	td->runstate >= TD_EXITED \|\| ss->attained)
4cd5b75b JA	213	continue;
	214
	215	td_iops = 0;
	216	td_bytes = 0;
	217	if (!td->o.group_reporting \|\|
	218	(td->o.group_reporting && td->groupid != prev_groupid)) {
	219	group_bw = 0;
	220	group_iops = 0;
	221	group_ramp_time_over = 0;
	222	}
	223	prev_groupid = td->groupid;
	224
	225	fio_gettime(&now, NULL);
	226	if (ss->ramp_time && !ss->ramp_time_over)
	227	/*
	228	* Begin recording data one second after ss->ramp_time
	229	* has elapsed
	230	*/
	231	if (utime_since(&td->epoch, &now) >= (ss->ramp_time + 1000000L))
	232	ss->ramp_time_over = 1;
	233
	234	td_io_u_lock(td);
	235	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
	236	td_iops += td->io_blocks[ddir];
	237	td_bytes += td->io_bytes[ddir];
	238	}
	239	td_io_u_unlock(td);
	240
	241	rate_time = mtime_since(&ss->prev_time, &now);
	242	memcpy(&ss->prev_time, &now, sizeof(now));
	243
	244	/*
	245	* Begin monitoring when job starts but don't actually use
	246	* data in checking stopping criterion until ss->ramp_time is
	247	* over. This ensures that we will have a sane value in
	248	* prev_iops/bw the first time through after ss->ramp_time
	249	* is done.
	250	*/
	251	if (ss->ramp_time_over) {
	252	group_bw += 1000 * (td_bytes - ss->prev_bytes) / rate_time;
	253	group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
	254	++group_ramp_time_over;
	255	}
	256	ss->prev_iops = td_iops;
	257	ss->prev_bytes = td_bytes;
	258
	259	if (td->o.group_reporting && !ss->last_in_group)
	260	continue;
	261
7be78a8a JA	262	/*
	263	* Don't begin checking criterion until ss->ramp_time is over
	264	* for at least one thread in group
	265	*/
4cd5b75b JA	266	if (!group_ramp_time_over)
	267	continue;
	268
7be78a8a JA	269	dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
	270	"groupid: %u, rate_msec: %ld, "
	271	"iops: %lu, bw: %lu, head: %d, tail: %d\n",
	272	i, td->groupid, rate_time, group_iops,
	273	group_bw, ss->head, ss->tail);
4cd5b75b JA	274
	275	if (steadystate_check_slope(&td->o))
	276	ret = steadystate_slope(group_iops, group_bw, td);
	277	else
	278	ret = steadystate_deviation(group_iops, group_bw, td);
	279
	280	if (ret) {
	281	if (td->o.group_reporting) {
	282	for_each_td(td2, j) {
	283	if (td2->groupid == td->groupid) {
	284	td2->ss.attained = 1;
	285	fio_mark_td_terminate(td2);
	286	}
	287	}
	288	} else {
	289	ss->attained = 1;
	290	fio_mark_td_terminate(td);
	291	}
	292	}
	293	}
	294	}
	295
	296