X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=stat.c;h=c2705d26f38bbd82f25b82882aa1f5910901a7d0;hp=7d93abd5c1a5bc2b5c11ed3b7181e4610f95e22c;hb=37f0c1ae23ad1716403d3d113c3dfdf41c47e329;hpb=2dc84ba74186eb9ef4dde0f08dc20571c62e001e

diff --git a/stat.c b/stat.c
index 7d93abd5..c2705d26 100644
--- a/stat.c
+++ b/stat.c
@@ -8,329 +8,159 @@
 #include <math.h>
 
 #include "fio.h"
+#include "diskutil.h"
+#include "ieee754.h"
 
-static struct itimerval itimer;
-static struct list_head disk_list = LIST_HEAD_INIT(disk_list);
-static dev_t last_dev;
-
-/*
- * Cheesy number->string conversion, complete with carry rounding error.
- */
-static char *num2str(unsigned long num, int maxlen, int base)
+void update_rusage_stat(struct thread_data *td)
 {
-	/*
-	 * could be passed in for 10^3 base, but every caller expects
-	 * 2^10 base right now.
-	 */
-	const unsigned int thousand = 1024;
-	char postfix[] = { 'K', 'M', 'G', 'P', 'E' };
-	char *buf;
-	int i;
-
-	buf = malloc(128);
-
-	for (i = 0; base > 1; i++)
-		base /= thousand;
-
-	do {
-		int len, carry = 0;
-
-		len = sprintf(buf, "%'lu", num);
-		if (len <= maxlen) {
-			buf[len] = postfix[i];
-			buf[len + 1] = '\0';
-			return buf;
-		}
+	struct thread_stat *ts = &td->ts;
 
-		if ((num % thousand) >= (thousand / 2))
-			carry = 1;
+	getrusage(RUSAGE_SELF, &td->ru_end);
 
-		num /= thousand;
-		num += carry;
-		i++;
-	} while (i <= 5);
+	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;
 
-	return buf;
+	memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
 }
 
-static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
+/*
+ * 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 in_flight;
-	char line[256];
-	FILE *f;
-	char *p;
+	unsigned int msb, error_bits, base, offset, idx;
 
-	f = fopen(du->path, "r");
-	if (!f)
-		return 1;
-
-	p = fgets(line, sizeof(line), f);
-	if (!p) {
-		fclose(f);
-		return 1;
-	}
-
-	if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
-		fclose(f);
-		return 1;
-	}
+	/* Find MSB starting from bit 0 */
+	if (val == 0)
+		msb = 0;
+	else
+		msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
 
-	fclose(f);
-	return 0;
-}
+	/*
+	 * 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;
 
-static void update_io_tick_disk(struct disk_util *du)
-{
-	struct disk_util_stat __dus, *dus, *ldus;
-	struct timeval t;
+	/* Compute the number of error bits to discard*/
+	error_bits = msb - FIO_IO_U_PLAT_BITS;
 
-	if (get_io_ticks(du, &__dus))
-		return;
+	/* Compute the number of buckets before the group */
+	base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
 
-	dus = &du->dus;
-	ldus = &du->last_dus;
+	/*
+	 * 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);
 
-	dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
-	dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
-	dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
-	dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
-	dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
-	dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
-	dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
-	dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
-	dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
-	dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
+	/* 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);
 
-	fio_gettime(&t, NULL);
-	du->msec += mtime_since(&du->time, &t);
-	memcpy(&du->time, &t, sizeof(t));
-	memcpy(ldus, &__dus, sizeof(__dus));
+	return idx;
 }
 
-void update_io_ticks(void)
+/*
+ * 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)
 {
-	struct list_head *entry;
-	struct disk_util *du;
+	unsigned int error_bits, k, base;
 
-	list_for_each(entry, &disk_list) {
-		du = list_entry(entry, struct disk_util, list);
-		update_io_tick_disk(du);
-	}
-}
+	assert(idx < FIO_IO_U_PLAT_NR);
 
-static int disk_util_exists(dev_t dev)
-{
-	struct list_head *entry;
-	struct disk_util *du;
+	/* 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;
 
-	list_for_each(entry, &disk_list) {
-		du = list_entry(entry, struct disk_util, list);
+	/* 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);
 
-		if (du->dev == dev)
-			return 1;
-	}
+	/* Find its bucket number of the group */
+	k = idx % FIO_IO_U_PLAT_VAL;
 
-	return 0;
+	/* Return the mean of the range of the bucket */
+	return base + ((k + 0.5) * (1 << error_bits));
 }
 
-static void disk_util_add(dev_t dev, char *path)
+static int double_cmp(const void *a, const void *b)
 {
-	struct disk_util *du, *__du;
-	struct list_head *entry;
-
-	du = malloc(sizeof(*du));
-	memset(du, 0, sizeof(*du));
-	INIT_LIST_HEAD(&du->list);
-	sprintf(du->path, "%s/stat", path);
-	du->name = strdup(basename(path));
-	du->dev = dev;
-
-	list_for_each(entry, &disk_list) {
-		__du = list_entry(entry, struct disk_util, list);
-
-		if (!strcmp(du->name, __du->name)) {
-			free(du->name);
-			free(du);
-			return;
-		}
-	}
+	const fio_fp64_t fa = *(const fio_fp64_t *) a;
+	const fio_fp64_t fb = *(const fio_fp64_t *) b;
+	int cmp = 0;
 
-	fio_gettime(&du->time, NULL);
-	get_io_ticks(du, &du->last_dus);
+	if (fa.u.f > fb.u.f)
+		cmp = 1;
+	else if (fa.u.f < fb.u.f)
+		cmp = -1;
 
-	list_add_tail(&du->list, &disk_list);
+	return cmp;
 }
 
-static int check_dev_match(dev_t dev, char *path)
+/*
+ * 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 int major, minor;
-	char line[256], *p;
-	FILE *f;
-
-	f = fopen(path, "r");
-	if (!f) {
-		perror("open path");
-		return 1;
-	}
-
-	p = fgets(line, sizeof(line), f);
-	if (!p) {
-		fclose(f);
-		return 1;
-	}
-
-	if (sscanf(p, "%u:%u", &major, &minor) != 2) {
-		fclose(f);
-		return 1;
-	}
+	unsigned long sum = 0;
+	unsigned int len, i, j = 0;
+	int is_last = 0;
 
-	if (((major << 8) | minor) == dev) {
-		fclose(f);
-		return 0;
-	}
+	len = 0;
+	while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
+		len++;
 
-	fclose(f);
-	return 1;
-}
+	/*
+	 * 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);
 
-static int find_block_dir(dev_t dev, char *path)
-{
-	struct dirent *dir;
-	struct stat st;
-	int found = 0;
-	DIR *D;
+	log_info("    clat percentiles (usec) :");
 
-	D = opendir(path);
-	if (!D)
-		return 0;
+	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)) {
+			assert(plist[j].u.f <= 100.0);
 
-	while ((dir = readdir(D)) != NULL) {
-		char full_path[256];
+			/* for formatting */
+			if (j != 0 && (j % 4) == 0)
+				log_info("                             ");
 
-		if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
-			continue;
+			/* end of the list */
+			is_last = (j == len - 1);
 
-		sprintf(full_path, "%s/%s", path, dir->d_name);
+			log_info(" %2.2fth=%u%c", plist[j], plat_idx_to_val(i),
+				 (is_last? '\n' : ','));
 
-		if (!strcmp(dir->d_name, "dev")) {
-			if (!check_dev_match(dev, full_path)) {
-				found = 1;
+			if (is_last)
 				break;
-			}
-		}
-
-		if (lstat(full_path, &st) == -1) {
-			perror("stat");
-			break;
-		}
-
-		if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
-			continue;
-
-		found = find_block_dir(dev, full_path);
-		if (found) {
-			strcpy(path, full_path);
-			break;
-		}
-	}
-
-	closedir(D);
-	return found;
-}
-
-void init_disk_util(struct thread_data *td)
-{
-	struct fio_file *f;
-	struct stat st;
-	char foo[PATH_MAX], tmp[PATH_MAX];
-	dev_t dev;
-	char *p;
-
-	if (!td->do_disk_util || (td->io_ops->flags & FIO_DISKLESSIO))
-		return;
-
-	/*
-	 * Just use the same file, they are on the same device.
-	 */
-	f = &td->files[0];
-	if (!stat(f->file_name, &st)) {
-		if (S_ISBLK(st.st_mode))
-			dev = st.st_rdev;
-		else
-			dev = st.st_dev;
-	} else {
-		/*
-		 * must be a file, open "." in that path
-		 */
-		strncpy(foo, f->file_name, PATH_MAX - 1);
-		p = dirname(foo);
-		if (stat(p, &st)) {
-			perror("disk util stat");
-			return;
-		}
 
-		dev = st.st_dev;
-	}
-
-	if (disk_util_exists(dev))
-		return;
-
-	/*
-	 * for an fs without a device, we will repeatedly stat through
-	 * sysfs which can take oodles of time for thousands of files. so
-	 * cache the last lookup and compare with that before going through
-	 * everything again.
-	 */
-	if (dev == last_dev)
-		return;
-
-	last_dev = dev;
-		
-	sprintf(foo, "/sys/block");
-	if (!find_block_dir(dev, foo))
-		return;
-
-	/*
-	 * If there's a ../queue/ directory there, we are inside a partition.
-	 * Check if that is the case and jump back. For loop/md/dm etc we
-	 * are already in the right spot.
-	 */
-	sprintf(tmp, "%s/../queue", foo);
-	if (!stat(tmp, &st)) {
-		p = dirname(foo);
-		sprintf(tmp, "%s/queue", p);
-		if (stat(tmp, &st)) {
-			log_err("unknown sysfs layout\n");
-			return;
+			if (j % 4 == 3)	/* for formatting */
+				log_info("\n");
+			j++;
 		}
-		strncpy(tmp, p, PATH_MAX - 1);
-		sprintf(foo, "%s", tmp);
 	}
-
-	if (td->ioscheduler)
-		td->sysfs_root = strdup(foo);
-
-	disk_util_add(dev, foo);
-}
-
-void disk_util_timer_arm(void)
-{
-	itimer.it_value.tv_sec = 0;
-	itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
-	setitimer(ITIMER_REAL, &itimer, NULL);
-}
-
-void update_rusage_stat(struct thread_data *td)
-{
-	struct thread_stat *ts = &td->ts;
-
-	getrusage(RUSAGE_SELF, &ts->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);
-	
-	memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
 }
 
 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
@@ -345,34 +175,39 @@ 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 = -1.0;
+		*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;
 
-	fprintf(f_out, "\nRun status group %d (all jobs):\n", id);
+	log_info("\nRun 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);
+
 		if (!rs->max_run[i])
 			continue;
 
-		p1 = num2str(rs->io_kb[i], 6, 1);
-		p2 = num2str(rs->agg[i], 6, 1);
-		p3 = num2str(rs->min_bw[i], 6, 1);
-		p4 = num2str(rs->max_bw[i], 6, 1);
+		p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p);
+		p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p);
+		p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p);
+		p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p);
 
-		fprintf(f_out, "%s: io=%siB, aggrb=%siB/s, minb=%siB/s, maxb=%siB/s, mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2, p3, p4, rs->min_run[0], rs->max_run[0]);
+		log_info("%s: io=%sB, aggrb=%sB/s, minb=%sB/s, maxb=%sB/s,"
+			 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2,
+						p3, p4, rs->min_run[i],
+						rs->max_run[i]);
 
 		free(p1);
 		free(p2);
@@ -381,88 +216,241 @@ static void show_group_stats(struct group_run_stats *rs, int id)
 	}
 }
 
-static void show_disk_util(void)
-{
-	struct disk_util_stat *dus;
-	struct list_head *entry, *next;
-	struct disk_util *du;
-	double util;
-
-	fprintf(f_out, "\nDisk stats (read/write):\n");
-
-	list_for_each(entry, &disk_list) {
-		du = list_entry(entry, struct disk_util, list);
-		dus = &du->dus;
+#define ts_total_io_u(ts)	\
+	((ts)->total_io_u[0] + (ts)->total_io_u[1])
 
-		util = (double) 100 * du->dus.io_ticks / (double) du->msec;
-		if (util > 100.0)
-			util = 100.0;
+static void stat_calc_dist(unsigned int *map, unsigned long total,
+			   double *io_u_dist)
+{
+	int i;
 
-		fprintf(f_out, "  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
+	/*
+	 * Do depth distribution calculations
+	 */
+	for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
+		if (total) {
+			io_u_dist[i] = (double) map[i] / (double) total;
+			io_u_dist[i] *= 100.0;
+			if (io_u_dist[i] < 0.1 && map[i])
+				io_u_dist[i] = 0.1;
+		} else
+			io_u_dist[i] = 0.0;
 	}
+}
+
+static void stat_calc_lat(struct thread_stat *ts, double *dst,
+			  unsigned int *src, int nr)
+{
+	unsigned long total = ts_total_io_u(ts);
+	int i;
 
 	/*
-	 * now free the list
+	 * Do latency distribution calculations
 	 */
-	list_for_each_safe(entry, next, &disk_list) {
-		list_del(entry);
-		du = list_entry(entry, struct disk_util, list);
-		free(du->name);
-		free(du);
+	for (i = 0; i < nr; i++) {
+		if (total) {
+			dst[i] = (double) src[i] / (double) total;
+			dst[i] *= 100.0;
+			if (dst[i] < 0.01 && src[i])
+				dst[i] = 0.01;
+		} else
+			dst[i] = 0.0;
 	}
 }
 
+static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
+{
+	stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
+}
+
+static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
+{
+	stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
+}
+
+static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean,
+			double *dev)
+{
+	if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
+		*min /= 1000;
+		*max /= 1000;
+		*mean /= 1000.0;
+		*dev /= 1000.0;
+		return 0;
+	}
+
+	return 1;
+}
+
 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 			     int ddir)
 {
 	const char *ddir_str[] = { "read ", "write" };
-	unsigned long min, max;
-	unsigned long long bw;
+	unsigned long min, max, runt;
+	unsigned long long bw, iops;
 	double mean, dev;
-	char *io_p, *bw_p;
+	char *io_p, *bw_p, *iops_p;
+	int i2p;
+
+	assert(ddir_rw(ddir));
 
 	if (!ts->runtime[ddir])
 		return;
 
-	bw = ts->io_bytes[ddir] / ts->runtime[ddir];
-	io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1);
-	bw_p = num2str(bw, 6, 1);
+	i2p = is_power_of_2(rs->kb_base);
+	runt = ts->runtime[ddir];
+
+	bw = (1000 * ts->io_bytes[ddir]) / runt;
+	io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
+	bw_p = num2str(bw, 6, 1, i2p);
 
-	fprintf(f_out, "  %s: io=%siB, bw=%siB/s, runt=%6lumsec\n", ddir_str[ddir], io_p, bw_p, ts->runtime[ddir]);
+	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=%6llumsec\n",
+					ddir_str[ddir], io_p, bw_p, iops_p,
+					ts->runtime[ddir]);
 
 	free(io_p);
 	free(bw_p);
+	free(iops_p);
 
-	if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
-		fprintf(f_out, "    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
+	if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
+		const char *base = "(usec)";
+		char *minp, *maxp;
 
-	if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
-		fprintf(f_out, "    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
+		if (!usec_to_msec(&min, &max, &mean, &dev))
+			base = "(msec)";
+
+		minp = num2str(min, 6, 1, 0);
+		maxp = num2str(max, 6, 1, 0);
+
+		log_info("    slat %s: min=%s, max=%s, avg=%5.02f,"
+			 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
+
+		free(minp);
+		free(maxp);
+	}
+	if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
+		const char *base = "(usec)";
+		char *minp, *maxp;
+
+		if (!usec_to_msec(&min, &max, &mean, &dev))
+			base = "(msec)";
+
+		minp = num2str(min, 6, 1, 0);
+		maxp = num2str(max, 6, 1, 0);
+
+		log_info("    clat %s: min=%s, max=%s, avg=%5.02f,"
+			 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
 
+		free(minp);
+		free(maxp);
+	}
+	if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
+		const char *base = "(usec)";
+		char *minp, *maxp;
+
+		if (!usec_to_msec(&min, &max, &mean, &dev))
+			base = "(msec)";
+
+		minp = num2str(min, 6, 1, 0);
+		maxp = num2str(max, 6, 1, 0);
+
+		log_info("     lat %s: min=%s, max=%s, avg=%5.02f,"
+			 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
+
+		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];
-		fprintf(f_out, "    bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg, mean, dev);
+		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);
+	}
+}
+
+static void show_lat(double *io_u_lat, int nr, const char **ranges,
+		     const char *msg)
+{
+	int new_line = 1, i, line = 0;
+
+	for (i = 0; i < nr; i++) {
+		if (io_u_lat[i] <= 0.0)
+			continue;
+		if (new_line) {
+			if (line)
+				log_info("\n");
+			log_info("     lat (%s): ", msg);
+			new_line = 0;
+			line = 0;
+		}
+		if (line)
+			log_info(", ");
+		log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
+		line++;
+		if (line == 5)
+			new_line = 1;
 	}
 }
 
-static void show_thread_status(struct thread_stat *ts,
-			       struct group_run_stats *rs)
+static void show_lat_u(double *io_u_lat_u)
+{
+	const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
+				 "250=", "500=", "750=", "1000=", };
+
+	show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
+}
+
+static void show_lat_m(double *io_u_lat_m)
+{
+	const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
+				 "250=", "500=", "750=", "1000=", "2000=",
+				 ">=2000=", };
+
+	show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
+}
+
+static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
+{
+	show_lat_u(io_u_lat_u);
+	log_info("\n");
+	show_lat_m(io_u_lat_m);
+	log_info("\n");
+}
+
+void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
 {
 	double usr_cpu, sys_cpu;
 	unsigned long runtime;
 	double io_u_dist[FIO_IO_U_MAP_NR];
-	double io_u_lat[FIO_IO_U_LAT_NR];
-	int i;
+	double io_u_lat_u[FIO_IO_U_LAT_U_NR];
+	double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 
-	if (!(ts->io_bytes[0] + ts->io_bytes[1]))
+	if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
+	    !(ts->total_io_u[0] + ts->total_io_u[1]))
 		return;
 
-	if (!ts->error)
-		fprintf(f_out, "%s: (groupid=%d): err=%2d: pid=%d\n", ts->name, ts->groupid, ts->error, ts->pid);
-	else
-		fprintf(f_out, "%s: (groupid=%d): err=%2d (%s): pid=%d\n", ts->name, ts->groupid, ts->error, ts->verror, ts->pid);
+	if (!ts->error) {
+		log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
+					ts->name, ts->groupid, ts->members,
+					ts->error, (int) ts->pid);
+	} else {
+		log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
+					ts->name, ts->groupid, ts->members,
+					ts->error, ts->verror, (int) ts->pid);
+	}
+
+	if (ts->description)
+		log_info("  Description  : [%s]\n", ts->description);
 
 	if (ts->io_bytes[DDIR_READ])
 		show_ddir_status(rs, ts, DDIR_READ);
@@ -480,31 +468,42 @@ static void show_thread_status(struct thread_stat *ts,
 		sys_cpu = 0;
 	}
 
-	fprintf(f_out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ts->ctx);
-
-	/*
-	 * Do depth distribution calculations
-	 */
-	for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
-		io_u_dist[i] = (double) ts->io_u_map[i] / (double) ts->total_io_u;
-		io_u_dist[i] *= 100.0;
+	log_info("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
+		 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
+
+	stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
+	log_info("  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
+		 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", 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]);
+
+	stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
+	log_info("     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
+		 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", 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]);
+	stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
+	log_info("     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
+		 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", 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]);
+	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->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);
+	if (ts->continue_on_error) {
+		log_info("     errors    : total=%lu, first_error=%d/<%s>\n",
+					ts->total_err_count,
+					ts->first_error,
+					strerror(ts->first_error));
 	}
-
-	fprintf(f_out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", 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]);
-
-	/*
-	 * Do latency distribution calculations
-	 */
-	for (i = 0; i < FIO_IO_U_LAT_NR; i++) {
-		io_u_lat[i] = (double) ts->io_u_lat[i] / (double) ts->total_io_u;
-		io_u_lat[i] *= 100.0;
-	}
-
-	fprintf(f_out, "     lat (msec): 2=%3.1f%%, 4=%3.1f%%, 10=%3.1f%%, 20=%3.1f%%, 50=%3.1f%%, 100=%3.1f%%\n", io_u_lat[0], io_u_lat[1], io_u_lat[2], io_u_lat[3], io_u_lat[4], io_u_lat[5]);
-	fprintf(f_out, "     lat (msec): 250=%3.1f%%, 500=%3.1f%%, 750=%3.1f%%, 1000=%3.1f%%, >=2000=%3.1f%%\n", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
-
-	if (ts->description)
-		fprintf(f_out, "%s\n", ts->description);
 }
 
 static void show_ddir_status_terse(struct thread_stat *ts,
@@ -514,42 +513,59 @@ 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];
 
-	fprintf(f_out, ",%llu,%llu,%lu", ts->io_bytes[ddir] >> 10, bw, ts->runtime[ddir]);
+	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))
-		fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
+		log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 	else
-		fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
+		log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 
 	if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
-		fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
+		log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
+	else
+		log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
+
+	if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
+		log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 	else
-		fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
+		log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 
 	if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 		double p_of_agg;
 
 		p_of_agg = mean * 100 / (double) rs->agg[ddir];
-		fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", min, max, p_of_agg, mean, dev);
+		log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 	} else
-		fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", 0UL, 0UL, 0.0, 0.0, 0.0);
+		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)
 {
+	double io_u_dist[FIO_IO_U_MAP_NR];
+	double io_u_lat_u[FIO_IO_U_LAT_U_NR];
+	double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 	double usr_cpu, sys_cpu;
+	int i;
 
-	fprintf(f_out, "%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;
 
@@ -560,37 +576,161 @@ static void show_thread_status_terse(struct thread_stat *ts,
 		sys_cpu = 0;
 	}
 
-	fprintf(f_out, ",%f%%,%f%%,%lu\n", usr_cpu, sys_cpu, ts->ctx);
+	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);
+
+	log_info("\n");
 }
 
-static void __sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
+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;
 }
 
-static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
+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)
 {
-	__sum_stat(&dst[DDIR_READ], &src[DDIR_READ], nr);
-	__sum_stat(&dst[DDIR_WRITE], &src[DDIR_WRITE], 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)
@@ -598,17 +738,13 @@ void show_run_stats(void)
 	struct group_run_stats *runstats, *rs;
 	struct thread_data *td;
 	struct thread_stat *threadstats, *ts;
-	int i, j, k, nr_ts, last_ts, members;
+	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
@@ -617,7 +753,7 @@ void show_run_stats(void)
 	nr_ts = 0;
 	last_ts = -1;
 	for_each_td(td, i) {
-		if (!td->group_reporting) {
+		if (!td->o.group_reporting) {
 			nr_ts++;
 			continue;
 		}
@@ -630,150 +766,146 @@ 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));
-		ts->clat_stat[0].min_val = -1UL;
-		ts->clat_stat[1].min_val = -1UL;
-		ts->slat_stat[0].min_val = -1UL;
-		ts->slat_stat[1].min_val = -1UL;
-		ts->bw_stat[0].min_val = -1UL;
-		ts->bw_stat[1].min_val = -1UL;
-	}
+	for (i = 0; i < nr_ts; i++)
+		init_thread_stat(&threadstats[i]);
 
 	j = 0;
 	last_ts = -1;
-	members = 0;
+	idx = 0;
 	for_each_td(td, i) {
+		if (idx && (!td->o.group_reporting ||
+		    (td->o.group_reporting && last_ts != td->groupid))) {
+			idx = 0;
+			j++;
+		}
+
+		last_ts = td->groupid;
+
 		ts = &threadstats[j];
 
-		members++;
+		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));
 
-		if (!ts->groupid) {
+		idx++;
+		ts->members++;
+
+		if (ts->groupid == -1) {
 			/*
 			 * These are per-group shared already
 			 */
-			ts->name = td->name;
-			ts->description = td->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;
 
 			/*
 			 * first pid in group, not very useful...
 			 */
 			ts->pid = td->pid;
-		}
-
-		if (td->error && !ts->error) {
-			ts->error = td->error;
-			ts->verror = td->verror;
-		}
-
-		sum_stat(ts->clat_stat, td->ts.clat_stat, members);
-		sum_stat(ts->slat_stat, td->ts.slat_stat, members);
-		sum_stat(ts->bw_stat, td->ts.bw_stat, members);
-
-		ts->stat_io_bytes[0] += td->ts.stat_io_bytes[0];
-		ts->stat_io_bytes[1] += td->ts.stat_io_bytes[1];
-
-		ts->usr_time += td->ts.usr_time;
-		ts->sys_time += td->ts.sys_time;
-		ts->ctx += td->ts.ctx;
-
-		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_LAT_NR; k++)
-			ts->io_u_lat[k] += td->ts.io_u_lat[k];
 
-		ts->total_io_u += td->ts.total_io_u;
-		ts->io_bytes[0] += td->ts.io_bytes[0];
-		ts->io_bytes[1] += td->ts.io_bytes[1];
-
-		if (ts->runtime[0] < td->ts.runtime[0])
-			ts->runtime[0] = td->ts.runtime[0];
-		if (ts->runtime[1] < td->ts.runtime[1])
-			ts->runtime[1] = td->ts.runtime[1];
-
-		ts->total_run_time += td->ts.total_run_time;
-
-		if (!td->group_reporting) {
-			members = 0;
-			j++;
-			continue;
+			ts->kb_base = td->o.kb_base;
+		} else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
+			log_info("fio: kb_base differs for jobs in group, using"
+				 " %u as the base\n", ts->kb_base);
+			kb_base_warned = 1;
 		}
-		if (last_ts == td->groupid)
-			continue;
 
-		if (last_ts != -1) {
-			members = 0;
-			j++;
+		ts->continue_on_error = td->o.continue_on_error;
+		ts->total_err_count += td->total_err_count;
+		ts->first_error = td->first_error;
+		if (!ts->error) {
+			if (!td->error && td->o.continue_on_error &&
+			    td->first_error) {
+				ts->error = td->first_error;
+				strcpy(ts->verror, td->verror);
+			} else  if (td->error) {
+				ts->error = td->error;
+				strcpy(ts->verror, td->verror);
+			}
 		}
 
-		last_ts = td->groupid;
+		sum_thread_stats(ts, &td->ts, idx);
 	}
 
 	for (i = 0; i < nr_ts; i++) {
-		unsigned long long rbw, wbw;
+		unsigned long long bw;
 
 		ts = &threadstats[i];
 		rs = &runstats[ts->groupid];
+		rs->kb_base = ts->kb_base;
+
+		for (j = 0; j <= DDIR_WRITE; j++) {
+			if (!ts->runtime[j])
+				continue;
+			if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
+				rs->min_run[j] = ts->runtime[j];
+			if (ts->runtime[j] > rs->max_run[j])
+				rs->max_run[j] = ts->runtime[j];
+
+			bw = 0;
+			if (ts->runtime[j]) {
+				unsigned long runt;
+
+				runt = ts->runtime[j];
+				bw = ts->io_bytes[j] / runt;
+			}
+			if (bw < rs->min_bw[j])
+				rs->min_bw[j] = bw;
+			if (bw > rs->max_bw[j])
+				rs->max_bw[j] = bw;
 
-		if (ts->runtime[0] < rs->min_run[0] || !rs->min_run[0])
-			rs->min_run[0] = ts->runtime[0];
-		if (ts->runtime[0] > rs->max_run[0])
-			rs->max_run[0] = ts->runtime[0];
-		if (ts->runtime[1] < rs->min_run[1] || !rs->min_run[1])
-			rs->min_run[1] = ts->runtime[1];
-		if (ts->runtime[1] > rs->max_run[1])
-			rs->max_run[1] = ts->runtime[1];
-
-		rbw = wbw = 0;
-		if (ts->runtime[0])
-			rbw = ts->io_bytes[0] / (unsigned long long) ts->runtime[0];
-		if (ts->runtime[1])
-			wbw = ts->io_bytes[1] / (unsigned long long) ts->runtime[1];
-
-		if (rbw < rs->min_bw[0])
-			rs->min_bw[0] = rbw;
-		if (wbw < rs->min_bw[1])
-			rs->min_bw[1] = wbw;
-		if (rbw > rs->max_bw[0])
-			rs->max_bw[0] = rbw;
-		if (wbw > rs->max_bw[1])
-			rs->max_bw[1] = wbw;
-
-		rs->io_kb[0] += ts->io_bytes[0] >> 10;
-		rs->io_kb[1] += ts->io_bytes[1] >> 10;
+			rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
+		}
 	}
 
 	for (i = 0; i < groupid + 1; i++) {
+		unsigned long max_run[2];
+
 		rs = &runstats[i];
+		max_run[0] = rs->max_run[0];
+		max_run[1] = rs->max_run[1];
 
 		if (rs->max_run[0])
-			rs->agg[0] = (rs->io_kb[0]*1024) / rs->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]*1024) / rs->max_run[1];
+			rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
 	}
 
 	/*
 	 * don't overwrite last signal output
 	 */
 	if (!terse_output)
-		printf("\n");
+		log_info("\n");
 
 	for (i = 0; i < nr_ts; i++) {
 		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();
 	}
@@ -785,24 +917,28 @@ void show_run_stats(void)
 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
 {
 	double val = data;
-	double delta, n;
+	double delta;
 
 	if (data > is->max_val)
 		is->max_val = data;
 	if (data < is->min_val)
 		is->min_val = data;
 
-	delta = val - is->mean;
-	n = is->samples + 1.0;
-	is->mean += delta / n;
-	is->S += delta * (val - is->mean);
+	delta = val - is->mean.u.f;
+	if (delta) {
+		is->mean.u.f += delta / (is->samples + 1.0);
+		is->S.u.f += delta * (val - is->mean.u.f);
+	}
 
 	is->samples++;
 }
 
 static void __add_log_sample(struct io_log *iolog, unsigned long val,
-			     enum fio_ddir ddir, unsigned long time)
+			     enum fio_ddir ddir, unsigned int bs,
+			     unsigned long t)
 {
+	const int nr_samples = iolog->nr_samples;
+
 	if (iolog->nr_samples == iolog->max_samples) {
 		int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
 
@@ -810,63 +946,134 @@ static void __add_log_sample(struct io_log *iolog, unsigned long val,
 		iolog->max_samples <<= 1;
 	}
 
-	iolog->log[iolog->nr_samples].val = val;
-	iolog->log[iolog->nr_samples].time = time;
-	iolog->log[iolog->nr_samples].ddir = ddir;
+	iolog->log[nr_samples].val = val;
+	iolog->log[nr_samples].time = t;
+	iolog->log[nr_samples].ddir = ddir;
+	iolog->log[nr_samples].bs = bs;
 	iolog->nr_samples++;
 }
 
 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
-			   unsigned long val, enum fio_ddir ddir)
+			   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)
 {
-	__add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
+	struct io_log *iolog;
+
+	if (!ddir_rw(ddir))
+		return;
+
+	iolog = agg_io_log[ddir];
+	__add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
 }
 
-void add_agg_sample(unsigned long val, enum fio_ddir ddir)
+static void add_clat_percentile_sample(struct thread_stat *ts,
+				unsigned long usec, enum fio_ddir ddir)
 {
-	struct io_log *iolog = agg_io_log[ddir];
+	unsigned int idx = plat_val_to_idx(usec);
+	assert(idx < FIO_IO_U_PLAT_NR);
 
-	__add_log_sample(iolog, val, ddir, mtime_since_genesis());
+	ts->io_u_plat[ddir][idx]++;
 }
 
 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
-		     unsigned long msec)
+		     unsigned long usec, unsigned int bs)
 {
 	struct thread_stat *ts = &td->ts;
 
-	add_stat_sample(&ts->clat_stat[ddir], msec);
+	if (!ddir_rw(ddir))
+		return;
+
+	add_stat_sample(&ts->clat_stat[ddir], usec);
+
+	if (td->clat_log)
+		add_log_sample(td, td->clat_log, usec, ddir, bs);
 
-	if (ts->clat_log)
-		add_log_sample(td, ts->clat_log, msec, ddir);
+	if (ts->clat_percentiles)
+		add_clat_percentile_sample(ts, usec, ddir);
 }
 
 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
-		     unsigned long msec)
+		     unsigned long usec, unsigned int bs)
+{
+	struct thread_stat *ts = &td->ts;
+
+	if (!ddir_rw(ddir))
+		return;
+
+	add_stat_sample(&ts->slat_stat[ddir], usec);
+
+	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,
+		    unsigned long usec, unsigned int bs)
 {
 	struct thread_stat *ts = &td->ts;
 
-	add_stat_sample(&ts->slat_stat[ddir], msec);
+	if (!ddir_rw(ddir))
+		return;
 
-	if (ts->slat_log)
-		add_log_sample(td, ts->slat_log, msec, ddir);
+	add_stat_sample(&ts->lat_stat[ddir], usec);
+
+	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,
+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;
 
-	if (spent < td->bw_avg_time)
+	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]) / spent;
+	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);
+	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];
 }