[fio.git] / steadystate.c

#include <stdlib.h>

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

bool steadystate = false;

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

	if (!steadystate)
		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->check_iops ? iops : bw;

	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];
				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;
		}

		ss->oldest_y = ss->check_iops ? ss->iops_data[ss->head] : 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);
		ss->criterion = ss->pct ? 100.0 * ss->slope / (ss->sum_y / ss->dur) : 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++)
				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;

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