X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=stat.c;h=e7195c2845c5a156ae441444e25e88776276b171;hp=0f259d273bd48e04e8f7752e522ff44bd72a1a3c;hb=250506294d3658d0eab0139d3829917a58b62c9a;hpb=02af09886db695e5ea2b7fd2a632733955f3c03f

diff --git a/stat.c b/stat.c
index 0f259d27..e7195c28 100644
--- a/stat.c
+++ b/stat.c
@@ -9,23 +9,169 @@
 
 #include "fio.h"
 #include "diskutil.h"
+#include "ieee754.h"
 
 void update_rusage_stat(struct thread_data *td)
 {
 	struct thread_stat *ts = &td->ts;
 
-	getrusage(RUSAGE_SELF, &ts->ru_end);
+	getrusage(RUSAGE_SELF, &td->ru_end);
 
-	ts->usr_time += mtime_since(&ts->ru_start.ru_utime,
-					&ts->ru_end.ru_utime);
-	ts->sys_time += mtime_since(&ts->ru_start.ru_stime,
-					&ts->ru_end.ru_stime);
-	ts->ctx += ts->ru_end.ru_nvcsw + ts->ru_end.ru_nivcsw
-			- (ts->ru_start.ru_nvcsw + ts->ru_start.ru_nivcsw);
-	ts->minf += ts->ru_end.ru_minflt - ts->ru_start.ru_minflt;
-	ts->majf += ts->ru_end.ru_majflt - ts->ru_start.ru_majflt;
+	ts->usr_time += mtime_since(&td->ru_start.ru_utime,
+					&td->ru_end.ru_utime);
+	ts->sys_time += mtime_since(&td->ru_start.ru_stime,
+					&td->ru_end.ru_stime);
+	ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
+			- (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
+	ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
+	ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
 
-	memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
+	memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
+}
+
+/*
+ * Given a latency, return the index of the corresponding bucket in
+ * the structure tracking percentiles.
+ *
+ * (1) find the group (and error bits) that the value (latency)
+ * belongs to by looking at its MSB. (2) find the bucket number in the
+ * group by looking at the index bits.
+ *
+ */
+static unsigned int plat_val_to_idx(unsigned int val)
+{
+	unsigned int msb, error_bits, base, offset, idx;
+
+	/* Find MSB starting from bit 0 */
+	if (val == 0)
+		msb = 0;
+	else
+		msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
+
+	/*
+	 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
+	 * all bits of the sample as index
+	 */
+	if (msb <= FIO_IO_U_PLAT_BITS)
+		return val;
+
+	/* Compute the number of error bits to discard*/
+	error_bits = msb - FIO_IO_U_PLAT_BITS;
+
+	/* Compute the number of buckets before the group */
+	base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
+
+	/*
+	 * Discard the error bits and apply the mask to find the
+         * index for the buckets in the group
+	 */
+	offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
+
+	/* Make sure the index does not exceed (array size - 1) */
+	idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1)?
+		(base + offset) : (FIO_IO_U_PLAT_NR - 1);
+
+	return idx;
+}
+
+/*
+ * Convert the given index of the bucket array to the value
+ * represented by the bucket
+ */
+static unsigned int plat_idx_to_val(unsigned int idx)
+{
+	unsigned int error_bits, k, base;
+
+	assert(idx < FIO_IO_U_PLAT_NR);
+
+	/* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
+	 * all bits of the sample as index */
+	if (idx < (FIO_IO_U_PLAT_VAL << 1) )
+		return idx;
+
+	/* Find the group and compute the minimum value of that group */
+	error_bits = (idx >> FIO_IO_U_PLAT_BITS) -1;
+	base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
+
+	/* Find its bucket number of the group */
+	k = idx % FIO_IO_U_PLAT_VAL;
+
+	/* Return the mean of the range of the bucket */
+	return base + ((k + 0.5) * (1 << error_bits));
+}
+
+static int double_cmp(const void *a, const void *b)
+{
+	const fio_fp64_t fa = *(const fio_fp64_t *) a;
+	const fio_fp64_t fb = *(const fio_fp64_t *) b;
+	int cmp = 0;
+
+	if (fa.u.f > fb.u.f)
+		cmp = 1;
+	else if (fa.u.f < fb.u.f)
+		cmp = -1;
+
+	return cmp;
+}
+
+/*
+ * Find and display the p-th percentile of clat
+ */
+static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
+				  fio_fp64_t *plist)
+{
+	unsigned long sum = 0;
+	unsigned int len, i, j = 0;
+	int is_last = 0;
+
+	len = 0;
+	while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
+		len++;
+
+	if (!len)
+		return;
+
+	/*
+	 * Sort the percentile list. Note that it may already be sorted if
+	 * we are using the default values, but since it's a short list this
+	 * isn't a worry. Also note that this does not work for NaN values.
+	 */
+	if (len > 1)
+		qsort((void*)plist, len, sizeof(plist[0]), double_cmp);
+
+	log_info("    clat percentiles (usec):\n     |");
+
+	for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
+		sum += io_u_plat[i];
+		while (sum >= (plist[j].u.f / 100.0 * nr)) {
+			char fbuf[8];
+
+			assert(plist[j].u.f <= 100.0);
+
+			/* for formatting */
+			if (j != 0 && (j % 4) == 0)
+				log_info("     |");
+
+			/* end of the list */
+			is_last = (j == len - 1);
+
+			if (plist[j].u.f < 10.0)
+				sprintf(fbuf, " %2.2f", plist[j].u.f);
+			else
+				sprintf(fbuf, "%2.2f", plist[j].u.f);
+
+			log_info(" %sth=[%5u]%c", fbuf, plat_idx_to_val(i),
+					is_last ? '\n' : ',');
+
+			if (is_last)
+				break;
+
+			if (j % 4 == 3)	/* for formatting */
+				log_info("\n");
+			if (++j == FIO_IO_U_LIST_MAX_LEN)
+				break;
+		}
+	}
 }
 
 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
@@ -40,23 +186,23 @@ static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 	*max = is->max_val;
 
 	n = (double) is->samples;
-	*mean = is->mean;
+	*mean = is->mean.u.f;
 
 	if (n > 1.0)
-		*dev = sqrt(is->S / (n - 1.0));
+		*dev = sqrt(is->S.u.f / (n - 1.0));
 	else
 		*dev = 0;
 
 	return 1;
 }
 
-static void show_group_stats(struct group_run_stats *rs, int id)
+void show_group_stats(struct group_run_stats *rs)
 {
 	char *p1, *p2, *p3, *p4;
 	const char *ddir_str[] = { "   READ", "  WRITE" };
 	int i;
 
-	log_info("\nRun status group %d (all jobs):\n", id);
+	log_info("Run status group %d (all jobs):\n", rs->groupid);
 
 	for (i = 0; i <= DDIR_WRITE; i++) {
 		const int i2p = is_power_of_2(rs->kb_base);
@@ -157,6 +303,8 @@ static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 	char *io_p, *bw_p, *iops_p;
 	int i2p;
 
+	assert(ddir_rw(ddir));
+
 	if (!ts->runtime[ddir])
 		return;
 
@@ -167,10 +315,10 @@ static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 	io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
 	bw_p = num2str(bw, 6, 1, i2p);
 
-	iops = (1000 * ts->total_io_u[ddir]) / runt;
+	iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 	iops_p = num2str(iops, 6, 1, 0);
 
-	log_info("  %s: io=%sB, bw=%sB/s, iops=%s, runt=%6lumsec\n",
+	log_info("  %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
 					ddir_str[ddir], io_p, bw_p, iops_p,
 					ts->runtime[ddir]);
 
@@ -226,11 +374,16 @@ static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 		free(minp);
 		free(maxp);
 	}
+	if (ts->clat_percentiles) {
+		show_clat_percentiles(ts->io_u_plat[ddir],
+					ts->clat_stat[ddir].samples,
+					ts->percentile_list);
+	}
 	if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 		double p_of_agg;
 
 		p_of_agg = mean * 100 / (double) rs->agg[ddir];
-		log_info("    bw (KB/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
+		log_info("     bw (KB/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
 			 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg,
 							mean, dev);
 	}
@@ -285,8 +438,7 @@ static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
 	log_info("\n");
 }
 
-static void show_thread_status(struct thread_stat *ts,
-			       struct group_run_stats *rs)
+void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
 {
 	double usr_cpu, sys_cpu;
 	unsigned long runtime;
@@ -349,9 +501,11 @@ static void show_thread_status(struct thread_stat *ts,
 					io_u_dist[1], io_u_dist[2],
 					io_u_dist[3], io_u_dist[4],
 					io_u_dist[5], io_u_dist[6]);
-	log_info("     issued r/w: total=%lu/%lu, short=%lu/%lu\n",
+	log_info("     issued r/w/d: total=%lu/%lu/%lu, short=%lu/%lu/%lu\n",
 					ts->total_io_u[0], ts->total_io_u[1],
-					ts->short_io_u[0], ts->short_io_u[1]);
+					ts->total_io_u[2],
+					ts->short_io_u[0], ts->short_io_u[1],
+					ts->short_io_u[2]);
 	stat_calc_lat_u(ts, io_u_lat_u);
 	stat_calc_lat_m(ts, io_u_lat_m);
 	show_latencies(io_u_lat_u, io_u_lat_m);
@@ -370,11 +524,13 @@ static void show_ddir_status_terse(struct thread_stat *ts,
 	unsigned long long bw;
 	double mean, dev;
 
+	assert(ddir_rw(ddir));
+
 	bw = 0;
 	if (ts->runtime[ddir])
 		bw = ts->io_bytes[ddir] / ts->runtime[ddir];
 
-	log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw,
+	log_info(";%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw,
 							ts->runtime[ddir]);
 
 	if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
@@ -401,6 +557,7 @@ static void show_ddir_status_terse(struct thread_stat *ts,
 		log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 }
 
+#define FIO_TERSE_VERSION	"2"
 
 static void show_thread_status_terse(struct thread_stat *ts,
 				     struct group_run_stats *rs)
@@ -411,11 +568,15 @@ static void show_thread_status_terse(struct thread_stat *ts,
 	double usr_cpu, sys_cpu;
 	int i;
 
-	log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
-
+	/* General Info */
+	log_info("%s;%s;%d;%d", FIO_TERSE_VERSION, ts->name, ts->groupid,
+				ts->error);
+	/* Log Read Status */
 	show_ddir_status_terse(ts, rs, 0);
+	/* Log Write Status */
 	show_ddir_status_terse(ts, rs, 1);
 
+	/* CPU Usage */
 	if (ts->total_run_time) {
 		double runt = (double) ts->total_run_time;
 
@@ -429,22 +590,28 @@ static void show_thread_status_terse(struct thread_stat *ts,
 	log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
 								ts->minf);
 
+	/* Calc % distribution of IO depths, usecond, msecond latency */
 	stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
 	stat_calc_lat_u(ts, io_u_lat_u);
 	stat_calc_lat_m(ts, io_u_lat_m);
 
+	/* Only show fixed 7 I/O depth levels*/
 	log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 			io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 			io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 
+	/* Microsecond latency */
 	for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 		log_info(";%3.2f%%", io_u_lat_u[i]);
+	/* Millisecond latency */
 	for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 		log_info(";%3.2f%%", io_u_lat_m[i]);
+	/* Additional output if continue_on_error set - default off*/
 	if (ts->continue_on_error)
 		log_info(";%lu;%d", ts->total_err_count, ts->first_error);
 	log_info("\n");
 
+	/* Additional output if description is set */
 	if (ts->description)
 		log_info(";%s", ts->description);
 
@@ -455,25 +622,126 @@ static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
 {
 	double mean, S;
 
+	if (src->samples == 0)
+		return;
+
 	dst->min_val = min(dst->min_val, src->min_val);
 	dst->max_val = max(dst->max_val, src->max_val);
-	dst->samples += src->samples;
 
 	/*
-	 * Needs a new method for calculating stddev, we cannot just
-	 * average them we do below for nr > 1
+	 * Compute new mean and S after the merge
+	 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
+	 *  #Parallel_algorithm>
 	 */
 	if (nr == 1) {
-		mean = src->mean;
-		S = src->S;
+		mean = src->mean.u.f;
+		S = src->S.u.f;
 	} else {
-		mean = ((src->mean * (double) (nr - 1))
-				+ dst->mean) / ((double) nr);
-		S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
+		double delta = src->mean.u.f - dst->mean.u.f;
+
+		mean = ((src->mean.u.f * src->samples) +
+			(dst->mean.u.f * dst->samples)) /
+			(dst->samples + src->samples);
+
+		S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
+			(dst->samples * src->samples) /
+			(dst->samples + src->samples);
 	}
 
-	dst->mean = mean;
-	dst->S = S;
+	dst->samples += src->samples;
+	dst->mean.u.f = mean;
+	dst->S.u.f = S;
+}
+
+void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
+{
+	int i;
+
+	for (i = 0; i < 2; i++) {
+		if (dst->max_run[i] < src->max_run[i])
+			dst->max_run[i] = src->max_run[i];
+		if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
+			dst->min_run[i] = src->min_run[i];
+		if (dst->max_bw[i] < src->max_bw[i])
+			dst->max_bw[i] = src->max_bw[i];
+		if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
+			dst->min_bw[i] = src->min_bw[i];
+
+		dst->io_kb[i] += src->io_kb[i];
+		dst->agg[i] += src->agg[i];
+	}
+
+}
+
+void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
+{
+	int l, k;
+
+	for (l = 0; l <= DDIR_WRITE; l++) {
+		sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
+		sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
+		sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
+		sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
+
+		dst->io_bytes[l] += src->io_bytes[l];
+
+		if (dst->runtime[l] < src->runtime[l])
+			dst->runtime[l] = src->runtime[l];
+	}
+
+	dst->usr_time += src->usr_time;
+	dst->sys_time += src->sys_time;
+	dst->ctx += src->ctx;
+	dst->majf += src->majf;
+	dst->minf += src->minf;
+
+	for (k = 0; k < FIO_IO_U_MAP_NR; k++)
+		dst->io_u_map[k] += src->io_u_map[k];
+	for (k = 0; k < FIO_IO_U_MAP_NR; k++)
+		dst->io_u_submit[k] += src->io_u_submit[k];
+	for (k = 0; k < FIO_IO_U_MAP_NR; k++)
+		dst->io_u_complete[k] += src->io_u_complete[k];
+	for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
+		dst->io_u_lat_u[k] += src->io_u_lat_u[k];
+	for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
+		dst->io_u_lat_m[k] += src->io_u_lat_m[k];
+
+	for (k = 0; k <= 2; k++) {
+		dst->total_io_u[k] += src->total_io_u[k];
+		dst->short_io_u[k] += src->short_io_u[k];
+	}
+
+	for (k = 0; k <= DDIR_WRITE; k++) {
+		int m;
+		for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
+			dst->io_u_plat[k][m] += src->io_u_plat[k][m];
+	}
+
+	dst->total_run_time += src->total_run_time;
+	dst->total_submit += src->total_submit;
+	dst->total_complete += src->total_complete;
+}
+
+void init_group_run_stat(struct group_run_stats *gs)
+{
+	memset(gs, 0, sizeof(*gs));
+	gs->min_bw[0] = gs->min_run[0] = ~0UL;
+	gs->min_bw[1] = gs->min_run[1] = ~0UL;
+}
+
+void init_thread_stat(struct thread_stat *ts)
+{
+	int j;
+
+	memset(ts, 0, sizeof(*ts));
+
+	for (j = 0; j <= DDIR_WRITE; j++) {
+		ts->lat_stat[j].min_val = -1UL;
+		ts->clat_stat[j].min_val = -1UL;
+		ts->slat_stat[j].min_val = -1UL;
+		ts->bw_stat[j].min_val = -1UL;
+	}
+	ts->groupid = -1;
 }
 
 void show_run_stats(void)
@@ -481,18 +749,13 @@ void show_run_stats(void)
 	struct group_run_stats *runstats, *rs;
 	struct thread_data *td;
 	struct thread_stat *threadstats, *ts;
-	int i, j, k, l, nr_ts, last_ts, idx;
+	int i, j, nr_ts, last_ts, idx;
 	int kb_base_warned = 0;
 
 	runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
 
-	for (i = 0; i < groupid + 1; i++) {
-		rs = &runstats[i];
-
-		memset(rs, 0, sizeof(*rs));
-		rs->min_bw[0] = rs->min_run[0] = ~0UL;
-		rs->min_bw[1] = rs->min_run[1] = ~0UL;
-	}
+	for (i = 0; i < groupid + 1; i++)
+		init_group_run_stat(&runstats[i]);
 
 	/*
 	 * find out how many threads stats we need. if group reporting isn't
@@ -514,18 +777,8 @@ void show_run_stats(void)
 
 	threadstats = malloc(nr_ts * sizeof(struct thread_stat));
 
-	for (i = 0; i < nr_ts; i++) {
-		ts = &threadstats[i];
-
-		memset(ts, 0, sizeof(*ts));
-		for (j = 0; j <= DDIR_WRITE; j++) {
-			ts->lat_stat[j].min_val = -1UL;
-			ts->clat_stat[j].min_val = -1UL;
-			ts->slat_stat[j].min_val = -1UL;
-			ts->bw_stat[j].min_val = -1UL;
-		}
-		ts->groupid = -1;
-	}
+	for (i = 0; i < nr_ts; i++)
+		init_thread_stat(&threadstats[i]);
 
 	j = 0;
 	last_ts = -1;
@@ -541,6 +794,12 @@ void show_run_stats(void)
 
 		ts = &threadstats[j];
 
+		ts->clat_percentiles = td->o.clat_percentiles;
+		if (td->o.overwrite_plist)
+			memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
+		else
+			memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
+
 		idx++;
 		ts->members++;
 
@@ -548,8 +807,13 @@ void show_run_stats(void)
 			/*
 			 * These are per-group shared already
 			 */
-			ts->name = td->o.name;
-			ts->description = td->o.description;
+			strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
+			if (td->o.description)
+				strncpy(ts->description, td->o.description,
+						FIO_JOBNAME_SIZE);
+			else
+				memset(ts->description, 0, FIO_JOBNAME_SIZE);
+
 			ts->groupid = td->groupid;
 
 			/*
@@ -571,52 +835,14 @@ void show_run_stats(void)
 			if (!td->error && td->o.continue_on_error &&
 			    td->first_error) {
 				ts->error = td->first_error;
-				ts->verror = td->verror;
+				strcpy(ts->verror, td->verror);
 			} else  if (td->error) {
 				ts->error = td->error;
-				ts->verror = td->verror;
+				strcpy(ts->verror, td->verror);
 			}
 		}
 
-		for (l = 0; l <= DDIR_WRITE; l++) {
-			sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
-			sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
-			sum_stat(&ts->lat_stat[l], &td->ts.lat_stat[l], idx);
-			sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
-
-			ts->stat_io_bytes[l] += td->ts.stat_io_bytes[l];
-			ts->io_bytes[l] += td->ts.io_bytes[l];
-
-			if (ts->runtime[l] < td->ts.runtime[l])
-				ts->runtime[l] = td->ts.runtime[l];
-		}
-
-		ts->usr_time += td->ts.usr_time;
-		ts->sys_time += td->ts.sys_time;
-		ts->ctx += td->ts.ctx;
-		ts->majf += td->ts.majf;
-		ts->minf += td->ts.minf;
-
-		for (k = 0; k < FIO_IO_U_MAP_NR; k++)
-			ts->io_u_map[k] += td->ts.io_u_map[k];
-		for (k = 0; k < FIO_IO_U_MAP_NR; k++)
-			ts->io_u_submit[k] += td->ts.io_u_submit[k];
-		for (k = 0; k < FIO_IO_U_MAP_NR; k++)
-			ts->io_u_complete[k] += td->ts.io_u_complete[k];
-		for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
-			ts->io_u_lat_u[k] += td->ts.io_u_lat_u[k];
-		for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
-			ts->io_u_lat_m[k] += td->ts.io_u_lat_m[k];
-
-
-		for (k = 0; k <= DDIR_WRITE; k++) {
-			ts->total_io_u[k] += td->ts.total_io_u[k];
-			ts->short_io_u[k] += td->ts.short_io_u[k];
-		}
-
-		ts->total_run_time += td->ts.total_run_time;
-		ts->total_submit += td->ts.total_submit;
-		ts->total_complete += td->ts.total_complete;
+		sum_thread_stats(ts, &td->ts, idx);
 	}
 
 	for (i = 0; i < nr_ts; i++) {
@@ -658,9 +884,9 @@ void show_run_stats(void)
 		max_run[1] = rs->max_run[1];
 
 		if (rs->max_run[0])
-			rs->agg[0] = (rs->io_kb[0] * rs->kb_base) / max_run[0];
+			rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
 		if (rs->max_run[1])
-			rs->agg[1] = (rs->io_kb[1] * rs->kb_base) / max_run[1];
+			rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
 	}
 
 	/*
@@ -673,15 +899,24 @@ void show_run_stats(void)
 		ts = &threadstats[i];
 		rs = &runstats[ts->groupid];
 
-		if (terse_output)
+		if (is_backend)
+			fio_server_send_ts(ts, rs);
+		else if (terse_output)
 			show_thread_status_terse(ts, rs);
 		else
 			show_thread_status(ts, rs);
 	}
 
 	if (!terse_output) {
-		for (i = 0; i < groupid + 1; i++)
-			show_group_stats(&runstats[i], i);
+		for (i = 0; i < groupid + 1; i++) {
+			rs = &runstats[i];
+
+			rs->groupid = i;
+			if (is_backend)
+				fio_server_send_gs(rs);
+			else
+				show_group_stats(rs);
+		}
 
 		show_disk_util();
 	}
@@ -700,10 +935,10 @@ static inline void add_stat_sample(struct io_stat *is, unsigned long data)
 	if (data < is->min_val)
 		is->min_val = data;
 
-	delta = val - is->mean;
+	delta = val - is->mean.u.f;
 	if (delta) {
-		is->mean += delta / (is->samples + 1.0);
-		is->S += delta * (val - is->mean);
+		is->mean.u.f += delta / (is->samples + 1.0);
+		is->S.u.f += delta * (val - is->mean.u.f);
 	}
 
 	is->samples++;
@@ -711,7 +946,7 @@ static inline void add_stat_sample(struct io_stat *is, unsigned long data)
 
 static void __add_log_sample(struct io_log *iolog, unsigned long val,
 			     enum fio_ddir ddir, unsigned int bs,
-			     unsigned long time)
+			     unsigned long t)
 {
 	const int nr_samples = iolog->nr_samples;
 
@@ -723,7 +958,7 @@ static void __add_log_sample(struct io_log *iolog, unsigned long val,
 	}
 
 	iolog->log[nr_samples].val = val;
-	iolog->log[nr_samples].time = time;
+	iolog->log[nr_samples].time = t;
 	iolog->log[nr_samples].ddir = ddir;
 	iolog->log[nr_samples].bs = bs;
 	iolog->nr_samples++;
@@ -733,25 +968,47 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
 			   unsigned long val, enum fio_ddir ddir,
 			   unsigned int bs)
 {
+	if (!ddir_rw(ddir))
+		return;
+
 	__add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch));
 }
 
 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
 {
-	struct io_log *iolog = agg_io_log[ddir];
+	struct io_log *iolog;
+
+	if (!ddir_rw(ddir))
+		return;
 
+	iolog = agg_io_log[ddir];
 	__add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
 }
 
+static void add_clat_percentile_sample(struct thread_stat *ts,
+				unsigned long usec, enum fio_ddir ddir)
+{
+	unsigned int idx = plat_val_to_idx(usec);
+	assert(idx < FIO_IO_U_PLAT_NR);
+
+	ts->io_u_plat[ddir][idx]++;
+}
+
 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
 		     unsigned long usec, unsigned int bs)
 {
 	struct thread_stat *ts = &td->ts;
 
+	if (!ddir_rw(ddir))
+		return;
+
 	add_stat_sample(&ts->clat_stat[ddir], usec);
 
-	if (ts->clat_log)
-		add_log_sample(td, ts->clat_log, usec, ddir, bs);
+	if (td->clat_log)
+		add_log_sample(td, td->clat_log, usec, ddir, bs);
+
+	if (ts->clat_percentiles)
+		add_clat_percentile_sample(ts, usec, ddir);
 }
 
 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
@@ -759,10 +1016,13 @@ void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
 {
 	struct thread_stat *ts = &td->ts;
 
+	if (!ddir_rw(ddir))
+		return;
+
 	add_stat_sample(&ts->slat_stat[ddir], usec);
 
-	if (ts->slat_log)
-		add_log_sample(td, ts->slat_log, usec, ddir, bs);
+	if (td->slat_log)
+		add_log_sample(td, td->slat_log, usec, ddir, bs);
 }
 
 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
@@ -770,29 +1030,61 @@ void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
 {
 	struct thread_stat *ts = &td->ts;
 
+	if (!ddir_rw(ddir))
+		return;
+
 	add_stat_sample(&ts->lat_stat[ddir], usec);
 
-	if (ts->lat_log)
-		add_log_sample(td, ts->lat_log, usec, ddir, bs);
+	if (td->lat_log)
+		add_log_sample(td, td->lat_log, usec, ddir, bs);
 }
 
 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
 		   struct timeval *t)
 {
 	struct thread_stat *ts = &td->ts;
-	unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
-	unsigned long rate;
+	unsigned long spent, rate;
 
+	if (!ddir_rw(ddir))
+		return;
+
+	spent = mtime_since(&td->bw_sample_time, t);
 	if (spent < td->o.bw_avg_time)
 		return;
 
-	rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) *
+	rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) *
 			1000 / spent / 1024;
 	add_stat_sample(&ts->bw_stat[ddir], rate);
 
-	if (ts->bw_log)
-		add_log_sample(td, ts->bw_log, rate, ddir, bs);
+	if (td->bw_log)
+		add_log_sample(td, td->bw_log, rate, ddir, bs);
+
+	fio_gettime(&td->bw_sample_time, NULL);
+	td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
+}
+
+void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
+		     struct timeval *t)
+{
+	struct thread_stat *ts = &td->ts;
+	unsigned long spent, iops;
+
+	if (!ddir_rw(ddir))
+		return;
+
+	spent = mtime_since(&td->iops_sample_time, t);
+	if (spent < td->o.iops_avg_time)
+		return;
+
+	iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent;
+
+	add_stat_sample(&ts->iops_stat[ddir], iops);
+
+	if (td->iops_log) {
+		assert(iops);
+		add_log_sample(td, td->iops_log, iops, ddir, 0);
+	}
 
-	fio_gettime(&ts->stat_sample_time[ddir], NULL);
-	ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
+	fio_gettime(&td->iops_sample_time, NULL);
+	td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
 }