Add completion percentiles to results output
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 3a785ee..40ca1d3 100644 (file)
--- a/stat.c
+++ b/stat.c
 #include <math.h>
 
 #include "fio.h"
+#include "diskutil.h"
+#include "lib/ieee754.h"
+
+void update_rusage_stat(struct thread_data *td)
+{
+       struct thread_stat *ts = &td->ts;
+
+       getrusage(RUSAGE_SELF, &td->ru_end);
+
+       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(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
+}
 
 /*
- * Cheesy number->string conversion, complete with carry rounding error.
+ * 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 char *num2str(unsigned long num, int maxlen, int base, int pow2)
+static unsigned int plat_val_to_idx(unsigned int val)
 {
-       char postfix[] = { ' ', 'K', 'M', 'G', 'P', 'E' };
-       unsigned int thousand;
-       char *buf;
-       int i;
+       unsigned int msb, error_bits, base, offset, idx;
 
-       if (pow2)
-               thousand = 1024;
+       /* Find MSB starting from bit 0 */
+       if (val == 0)
+               msb = 0;
        else
-               thousand = 1000;
+               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;
+}
 
-       buf = malloc(128);
+/*
+ * 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;
+}
+
+unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
+                                  fio_fp64_t *plist, unsigned int **output,
+                                  unsigned int *maxv, unsigned int *minv)
+{
+       unsigned long sum = 0;
+       unsigned int len, i, j = 0;
+       unsigned int oval_len = 0;
+       unsigned int *ovals = NULL;
+       int is_last;
+
+       *minv = -1U;
+       *maxv = 0;
 
-       for (i = 0; base > 1; i++)
-               base /= thousand;
+       len = 0;
+       while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
+               len++;
 
-       do {
-               int len, carry = 0;
+       if (!len)
+               return 0;
 
-               len = sprintf(buf, "%'lu", num);
-               if (len <= maxlen) {
-                       if (i >= 1) {
-                               buf[len] = postfix[i];
-                               buf[len + 1] = '\0';
+       /*
+        * 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);
+
+       /*
+        * Calculate bucket values, note down max and min values
+        */
+       is_last = 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);
+
+                       if (j == oval_len) {
+                               oval_len += 100;
+                               ovals = realloc(ovals, oval_len * sizeof(unsigned int));
                        }
-                       return buf;
-               }
 
-               if ((num % thousand) >= (thousand / 2))
-                       carry = 1;
+                       ovals[j] = plat_idx_to_val(i);
+                       if (ovals[j] < *minv)
+                               *minv = ovals[j];
+                       if (ovals[j] > *maxv)
+                               *maxv = ovals[j];
+
+                       is_last = (j == len - 1);
+                       if (is_last)
+                               break;
 
-               num /= thousand;
-               num += carry;
-               i++;
-       } while (i <= 5);
+                       j++;
+               }
+       }
 
-       return buf;
+       *output = ovals;
+       return len;
 }
 
-void update_rusage_stat(struct thread_data *td)
+/*
+ * 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)
 {
-       struct thread_stat *ts = &td->ts;
+       unsigned int len, j = 0, minv, maxv;
+       unsigned int *ovals;
+       int is_last, scale_down;
+
+       len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
+       if (!len)
+               goto out;
+
+       /*
+        * We default to usecs, but if the value range is such that we
+        * should scale down to msecs, do that.
+        */
+       if (minv > 2000 && maxv > 99999) {
+               scale_down = 1;
+               log_info("    clat percentiles (msec):\n     |");
+       } else {
+               scale_down = 0;
+               log_info("    clat percentiles (usec):\n     |");
+       }
+
+       for (j = 0; j < len; j++) {
+               char fbuf[8];
+
+               /* for formatting */
+               if (j != 0 && (j % 4) == 0)
+                       log_info("     |");
 
-       getrusage(RUSAGE_SELF, &ts->ru_end);
+               /* end of the list */
+               is_last = (j == len - 1);
 
-       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));
+               if (plist[j].u.f < 10.0)
+                       sprintf(fbuf, " %2.2f", plist[j].u.f);
+               else
+                       sprintf(fbuf, "%2.2f", plist[j].u.f);
+
+               if (scale_down)
+                       ovals[j] = (ovals[j] + 999) / 1000;
+
+               log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
+
+               if (is_last)
+                       break;
+
+               if (j % 4 == 3) /* for formatting */
+                       log_info("\n");
+       }
+
+out:
+       if (ovals)
+               free(ovals);
 }
 
-static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
-                   double *mean, double *dev)
+int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
+            double *mean, double *dev)
 {
        double n = is->samples;
 
@@ -77,34 +244,39 @@ 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;
 
-       log_info("\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, 1000, 1);
-               p2 = num2str(rs->agg[i], 6, 1000, 1);
-               p3 = num2str(rs->min_bw[i], 6, 1000, 1);
-               p4 = num2str(rs->max_bw[i], 6, 1000, 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);
 
-               log_info("%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[i], rs->max_run[i]);
+               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);
@@ -116,7 +288,8 @@ static void show_group_stats(struct group_run_stats *rs, int id)
 #define ts_total_io_u(ts)      \
        ((ts)->total_io_u[0] + (ts)->total_io_u[1])
 
-static void stat_calc_dist(struct thread_stat *ts, double *io_u_dist)
+static void stat_calc_dist(unsigned int *map, unsigned long total,
+                          double *io_u_dist)
 {
        int i;
 
@@ -124,83 +297,214 @@ static void stat_calc_dist(struct thread_stat *ts, double *io_u_dist)
         * 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(ts);
-               io_u_dist[i] *= 100.0;
-               if (io_u_dist[i] < 0.1 && ts->io_u_map[i])
-                       io_u_dist[i] = 0.1;
+               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 *io_u_lat)
+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;
 
        /*
         * 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(ts);
-               io_u_lat[i] *= 100.0;
-               if (io_u_lat[i] < 0.01 && ts->io_u_lat[i])
-                       io_u_lat[i] = 0.01;
+       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 void display_lat(const char *name, unsigned long min, unsigned long max,
+                       double mean, double 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("    %s %s: min=%s, max=%s, avg=%5.02f,"
+                " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
+
+       free(minp);
+       free(maxp);
+}
+
 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 min, max, runt;
        unsigned long long bw, iops;
        double mean, dev;
        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];
-       iops = (1000 * ts->total_io_u[ddir]) / ts->runtime[ddir];
-       io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1000, 1);
-       bw_p = num2str(bw, 6, 1000, 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);
+
+       iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
        iops_p = num2str(iops, 6, 1, 0);
 
-       log_info("  %s: io=%siB, bw=%siB/s, iops=%s, runt=%6lumsec\n", ddir_str[ddir], io_p, bw_p, iops_p, ts->runtime[ddir]);
+       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))
-               log_info("    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
-
+               display_lat("slat", min, max, mean, dev);
        if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
-               log_info("    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
-
+               display_lat("clat", min, max, mean, dev);
+       if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
+               display_lat(" lat", min, max, mean, dev);
+
+       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;
+               double p_of_agg = 100.0;
+               const char *bw_str = "KB";
+
+               if (rs->agg[ddir]) {
+                       p_of_agg = mean * 100 / (double) rs->agg[ddir];
+                       if (p_of_agg > 100.0)
+                               p_of_agg = 100.0;
+               }
+
+               if (mean > 999999.9) {
+                       min /= 1000.0;
+                       max /= 1000.0;
+                       mean /= 1000.0;
+                       dev /= 1000.0;
+                       bw_str = "MB";
+               }
 
-               p_of_agg = mean * 100 / (double) rs->agg[ddir];
-               log_info("    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 (%s/s)  : min=%5lu, max=%5lu, per=%3.2f%%,"
+                        " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
+                                                       p_of_agg, mean, dev);
        }
 }
 
-static void show_thread_status(struct thread_stat *ts,
-                              struct group_run_stats *rs)
+static int show_lat(double *io_u_lat, int nr, const char **ranges,
+                   const char *msg)
+{
+       int new_line = 1, i, line = 0, shown = 0;
+
+       for (i = 0; i < nr; i++) {
+               if (io_u_lat[i] <= 0.0)
+                       continue;
+               shown = 1;
+               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;
+       }
+
+       if (shown)
+               log_info("\n");
+
+       return shown;
+}
+
+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(struct thread_stat *ts, double *io_u_lat_u,
+                          double *io_u_lat_m)
+{
+       stat_calc_lat_u(ts, io_u_lat_u);
+       stat_calc_lat_m(ts, io_u_lat_m);
+
+       show_lat_u(io_u_lat_u);
+       show_lat_m(io_u_lat_m);
+}
+
+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];
+       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)
-               log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n", ts->name, ts->groupid, ts->members, ts->error, 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, 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)
+       if (strlen(ts->description))
                log_info("  Description  : [%s]\n", ts->description);
 
        if (ts->io_bytes[DDIR_READ])
@@ -208,6 +512,8 @@ static void show_thread_status(struct thread_stat *ts,
        if (ts->io_bytes[DDIR_WRITE])
                show_ddir_status(rs, ts, DDIR_WRITE);
 
+       show_latencies(ts, io_u_lat_u, io_u_lat_m);
+
        runtime = ts->total_run_time;
        if (runtime) {
                double runt = (double) runtime;
@@ -219,29 +525,64 @@ static void show_thread_status(struct thread_stat *ts,
                sys_cpu = 0;
        }
 
-       log_info("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ts->ctx);
-
-       stat_calc_dist(ts, io_u_dist);
-       stat_calc_lat(ts, io_u_lat);
-
-       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]);
-
-       log_info("     lat (msec): 2=%3.2f%%, 4=%3.2f%%, 10=%3.2f%%, 20=%3.2f%%, 50=%3.2f%%, 100=%3.2f%%\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]);
-       log_info("     lat (msec): 250=%3.2f%%, 500=%3.2f%%, 750=%3.2f%%, 1000=%3.2f%%, >=2000=%3.2f%%\n", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
+       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    : total=r=%lu/w=%lu/d=%lu,"
+                                " short=r=%lu/w=%lu/d=%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]);
+       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));
+       }
 }
 
 static void show_ddir_status_terse(struct thread_stat *ts,
                                   struct group_run_stats *rs, int ddir)
 {
        unsigned long min, max;
-       unsigned long long bw;
+       unsigned long long bw, iops;
+       unsigned int *ovals = NULL;
        double mean, dev;
+       unsigned int len, minv, maxv;
+       int i;
+
+       assert(ddir_rw(ddir));
 
-       bw = 0;
-       if (ts->runtime[ddir])
-               bw = ts->io_bytes[ddir] / ts->runtime[ddir];
+       iops = bw = 0;
+       if (ts->runtime[ddir]) {
+               uint64_t runt = ts->runtime[ddir];
 
-       log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw, ts->runtime[ddir]);
+               bw = ts->io_bytes[ddir] / runt;
+               iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
+       }
+
+       log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
+                                                       ts->runtime[ddir]);
 
        if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
                log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
@@ -253,28 +594,122 @@ static void show_ddir_status_terse(struct thread_stat *ts,
        else
                log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 
+       if (ts->clat_percentiles) {
+               len = calc_clat_percentiles(ts->io_u_plat[ddir],
+                                       ts->clat_stat[ddir].samples,
+                                       ts->percentile_list, &ovals, &maxv,
+                                       &minv);
+       } else
+               len = 0;
+
+       for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
+               if (i >= len) {
+                       log_info(";0%%=0");
+                       continue;
+               }
+               log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
+       }
+
+       if (calc_lat(&ts->lat_stat[ddir], &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 (ovals)
+               free(ovals);
+
        if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
-               double p_of_agg;
+               double p_of_agg = 100.0;
+
+               if (rs->agg[ddir]) {
+                       p_of_agg = mean * 100 / (double) rs->agg[ddir];
+                       if (p_of_agg > 100.0)
+                               p_of_agg = 100.0;
+               }
 
-               p_of_agg = mean * 100 / (double) rs->agg[ddir];
                log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
        } else
                log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 }
 
-
-static void show_thread_status_terse(struct thread_stat *ts,
-                                    struct group_run_stats *rs)
+static void show_thread_status_terse_v2(struct thread_stat *ts,
+                                       struct group_run_stats *rs)
 {
        double io_u_dist[FIO_IO_U_MAP_NR];
-       double io_u_lat[FIO_IO_U_LAT_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;
 
-       log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
+       /* General Info */
+       log_info("2;%s;%d;%d", 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;
+
+               usr_cpu = (double) ts->usr_time * 100 / runt;
+               sys_cpu = (double) ts->sys_time * 100 / runt;
+       } else {
+               usr_cpu = 0;
+               sys_cpu = 0;
+       }
+
+       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");
+}
+
+#define FIO_TERSE_VERSION      "3"
+
+static void show_thread_status_terse_v3(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;
+
+       /* General Info */
+       log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
+                                       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;
 
@@ -285,44 +720,174 @@ static void show_thread_status_terse(struct thread_stat *ts,
                sys_cpu = 0;
        }
 
-       log_info(";%f%%;%f%%;%lu", usr_cpu, sys_cpu, ts->ctx);
+       log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
+                                                               ts->minf);
 
-       stat_calc_dist(ts, io_u_dist);
-       stat_calc_lat(ts, io_u_lat);
+       /* 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);
 
-       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]);
+       /* 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]);
 
-       log_info(";%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%\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]);
-       log_info(";%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
+       /* 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]);
 
-       if (ts->description)
-               log_info(";%s", ts->description);
+       /* disk util stats, if any */
+       show_disk_util(1);
 
+       /* 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 (strlen(ts->description))
+               log_info(";%s", ts->description);
+}
+
+static void show_thread_status_terse(struct thread_stat *ts,
+                                    struct group_run_stats *rs)
+{
+       if (terse_version == 2)
+               show_thread_status_terse_v2(ts, rs);
+       else if (terse_version == 3)
+               show_thread_status_terse_v3(ts, rs);
+       else
+               log_err("fio: bad terse version!? %d\n", terse_version);
 }
 
 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->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->mean = mean;
-       dst->S = S;
+       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)
@@ -330,17 +895,13 @@ 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
@@ -362,17 +923,8 @@ 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->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;
@@ -388,6 +940,12 @@ 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++;
 
@@ -395,46 +953,42 @@ 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;
 
                        /*
                         * first pid in group, not very useful...
                         */
                        ts->pid = td->pid;
-               }
 
-               if (td->error && !ts->error) {
-                       ts->error = td->error;
-                       ts->verror = td->verror;
+                       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;
                }
 
-               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->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->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);
+                       }
                }
 
-               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];
-
-               for (k = 0; k <= DDIR_WRITE; k++)
-                       ts->total_io_u[k] += td->ts.total_io_u[k];
-
-               ts->total_run_time += td->ts.total_run_time;
+               sum_thread_stats(ts, &td->ts, idx);
        }
 
        for (i = 0; i < nr_ts; i++) {
@@ -442,6 +996,7 @@ void show_run_stats(void)
 
                ts = &threadstats[i];
                rs = &runstats[ts->groupid];
+               rs->kb_base = ts->kb_base;
 
                for (j = 0; j <= DDIR_WRITE; j++) {
                        if (!ts->runtime[j])
@@ -452,49 +1007,69 @@ void show_run_stats(void)
                                rs->max_run[j] = ts->runtime[j];
 
                        bw = 0;
-                       if (ts->runtime[j])
-                               bw = ts->io_bytes[j] / (unsigned long long) ts->runtime[j];
+                       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;
 
-                       rs->io_kb[j] += ts->io_bytes[j] >> 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];
 
-               show_disk_util();
+               rs->groupid = i;
+               if (is_backend)
+                       fio_server_send_gs(rs);
+               else if (!terse_output)
+                       show_group_stats(rs);
        }
 
+       if (is_backend)
+               fio_server_send_du();
+       else if (!terse_output)
+               show_disk_util(0);
+
+       free_disk_util();
+
        free(runstats);
        free(threadstats);
 }
@@ -509,18 +1084,24 @@ 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++;
 }
 
 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->avg_last = t;
+
        if (iolog->nr_samples == iolog->max_samples) {
                int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
 
@@ -528,63 +1109,206 @@ 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 inline void reset_io_stat(struct io_stat *ios)
+{
+       ios->max_val = ios->min_val = ios->samples = 0;
+       ios->mean.u.f = ios->S.u.f = 0;
+}
+
 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)
 {
-       __add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
+       unsigned long elapsed, this_window;
+
+       if (!ddir_rw(ddir))
+               return;
+
+       elapsed = mtime_since_now(&td->epoch);
+
+       /*
+        * If no time averaging, just add the log sample.
+        */
+       if (!iolog->avg_msec) {
+               __add_log_sample(iolog, val, ddir, bs, elapsed);
+               return;
+       }
+
+       /*
+        * Add the sample. If the time period has passed, then
+        * add that entry to the log and clear.
+        */
+       add_stat_sample(&iolog->avg_window[ddir], val);
+
+       /*
+        * If period hasn't passed, adding the above sample is all we
+        * need to do.
+        */
+       this_window = elapsed - iolog->avg_last;
+       if (this_window < iolog->avg_msec)
+               return;
+
+       /*
+        * Note an entry in the log. Use the mean from the logged samples,
+        * making sure to properly round up. Only write a log entry if we
+        * had actual samples done.
+        */
+       if (iolog->avg_window[DDIR_READ].samples) {
+               unsigned long mr;
+
+               mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
+               __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
+       }
+       if (iolog->avg_window[DDIR_WRITE].samples) {
+               unsigned long mw;
+
+               mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
+               __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
+       }
+
+       reset_io_stat(&iolog->avg_window[DDIR_READ]);
+       reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
+       iolog->avg_last = elapsed;
+}
+
+void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
+{
+       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 (ts->clat_log)
-               add_log_sample(td, ts->clat_log, msec, ddir);
+       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,
-                    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;
 
+       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;
-       add_stat_sample(&ts->bw_stat[ddir], rate);
+       /*
+        * Compute both read and write rates for the interval.
+        */
+       for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
+               uint64_t delta;
+
+               delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
+               if (!delta)
+                       continue; /* No entries for interval */
+
+               rate = delta * 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);
+
+               td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
+       }
+
+       fio_gettime(&td->bw_sample_time, NULL);
+}
+
+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;
+
+       /*
+        * Compute both read and write rates for the interval.
+        */
+       for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
+               uint64_t delta;
+
+               delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
+               if (!delta)
+                       continue; /* No entries for interval */
+
+               iops = (delta * 1000) / spent;
+               add_stat_sample(&ts->iops_stat[ddir], iops);
+
+               if (td->iops_log)
+                       add_log_sample(td, td->iops_log, iops, ddir, 0);
+
+               td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
+       }
 
-       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);
 }