Merge branch 'master' into gfio
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 48e9d76..402a73f 100644 (file)
--- a/stat.c
+++ b/stat.c
@@ -9,14 +9,16 @@
 
 #include "fio.h"
 #include "diskutil.h"
-#include "ieee754.h"
+#include "lib/ieee754.h"
+#include "json.h"
+#include "lib/getrusage.h"
+#include "idletime.h"
 
 void update_rusage_stat(struct thread_data *td)
 {
        struct thread_stat *ts = &td->ts;
 
-       getrusage(RUSAGE_SELF, &td->ru_end);
-
+       fio_getrusage(&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,
@@ -63,12 +65,12 @@ static unsigned int plat_val_to_idx(unsigned int val)
 
        /*
         * Discard the error bits and apply the mask to find the
-         * index for the buckets in the group
+        * 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)?
+       idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
                (base + offset) : (FIO_IO_U_PLAT_NR - 1);
 
        return idx;
@@ -86,11 +88,11 @@ static unsigned int plat_idx_to_val(unsigned int idx)
 
        /* 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) )
+       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;
+       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 */
@@ -114,22 +116,25 @@ static int double_cmp(const void *a, const void *b)
        return cmp;
 }
 
-/*
- * Find and display the p-th percentile of clat
- */
-static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
-                                 fio_fp64_t *plist)
+unsigned 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;
-       int is_last = 0;
+       unsigned int oval_len = 0;
+       unsigned int *ovals = NULL;
+       int is_last;
+
+       *minv = -1U;
+       *maxv = 0;
 
        len = 0;
        while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
                len++;
 
        if (!len)
-               return;
+               return 0;
 
        /*
         * Sort the percentile list. Note that it may already be sorted if
@@ -137,45 +142,104 @@ static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
         * isn't a worry. Also note that this does not work for NaN values.
         */
        if (len > 1)
-               qsort((void*)plist, len, sizeof(plist[0]), double_cmp);
-
-       log_info("    clat percentiles (usec):\n     |");
+               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)) {
-                       char fbuf[8];
-
                        assert(plist[j].u.f <= 100.0);
 
-                       /* for formatting */
-                       if (j != 0 && (j % 4) == 0)
-                               log_info("     |");
-
-                       /* end of the list */
-                       is_last = (j == len - 1);
-
-                       if (plist[j].u.f < 10.0)
-                               sprintf(fbuf, " %2.2f", plist[j].u.f);
-                       else
-                               sprintf(fbuf, "%2.2f", plist[j].u.f);
+                       if (j == oval_len) {
+                               oval_len += 100;
+                               ovals = realloc(ovals, oval_len * sizeof(unsigned int));
+                       }
 
-                       log_info(" %sth=[%5u]%c", fbuf, plat_idx_to_val(i),
-                                       is_last ? '\n' : ',');
+                       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;
 
-                       if (j % 4 == 3) /* for formatting */
-                               log_info("\n");
-                       if (++j == FIO_IO_U_LIST_MAX_LEN)
-                               break;
+                       j++;
                }
        }
+
+       *output = ovals;
+       return len;
 }
 
-static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
-                   double *mean, double *dev)
+/*
+ * 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 precision)
+{
+       unsigned int len, j = 0, minv, maxv;
+       unsigned int *ovals;
+       int is_last, per_line, scale_down;
+       char fmt[32];
+
+       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     |");
+       }
+
+       snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
+       per_line = (80 - 7) / (precision + 14);
+
+       for (j = 0; j < len; j++) {
+               char fbuf[16], *ptr = fbuf;
+
+               /* for formatting */
+               if (j != 0 && (j % per_line) == 0)
+                       log_info("     |");
+
+               /* end of the list */
+               is_last = (j == len - 1);
+
+               if (plist[j].u.f < 10.0)
+                       ptr += sprintf(fbuf, " ");
+
+               snprintf(ptr, sizeof(fbuf), fmt, 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 % per_line) == per_line - 1)     /* for formatting */
+                       log_info("\n");
+       }
+
+out:
+       if (ovals)
+               free(ovals);
+}
+
+int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
+            double *mean, double *dev)
 {
        double n = is->samples;
 
@@ -199,12 +263,12 @@ static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 void show_group_stats(struct group_run_stats *rs)
 {
        char *p1, *p2, *p3, *p4;
-       const char *ddir_str[] = { "   READ", "  WRITE" };
+       const char *ddir_str[] = { "   READ", "  WRITE" , "   TRIM"};
        int i;
 
        log_info("\nRun status group %d (all jobs):\n", rs->groupid);
 
-       for (i = 0; i <= DDIR_WRITE; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                const int i2p = is_power_of_2(rs->kb_base);
 
                if (!rs->max_run[i])
@@ -216,9 +280,9 @@ void show_group_stats(struct group_run_stats *rs)
                p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p);
 
                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]);
+                        " mint=%llumsec, maxt=%llumsec\n",
+                               rs->unified_rw_rep ? "  MIXED" : ddir_str[i],
+                               p1, p2, p3, p4, rs->min_run[i], rs->max_run[i]);
 
                free(p1);
                free(p2);
@@ -227,11 +291,7 @@ void show_group_stats(struct group_run_stats *rs)
        }
 }
 
-#define ts_total_io_u(ts)      \
-       ((ts)->total_io_u[0] + (ts)->total_io_u[1])
-
-static void stat_calc_dist(unsigned int *map, unsigned long total,
-                          double *io_u_dist)
+void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 {
        int i;
 
@@ -252,7 +312,7 @@ static void stat_calc_dist(unsigned int *map, unsigned long total,
 static void stat_calc_lat(struct thread_stat *ts, double *dst,
                          unsigned int *src, int nr)
 {
-       unsigned long total = ts_total_io_u(ts);
+       unsigned long total = ddir_rw_sum(ts->total_io_u);
        int i;
 
        /*
@@ -269,34 +329,39 @@ static void stat_calc_lat(struct thread_stat *ts, double *dst,
        }
 }
 
-static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
+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)
+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)
+static void display_lat(const char *name, 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;
-       }
+       const char *base = "(usec)";
+       char *minp, *maxp;
 
-       return 1;
+       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" };
+       const char *ddir_str[] = { "read ", "write", "trim" };
        unsigned long min, max, runt;
        unsigned long long bw, iops;
        double mean, dev;
@@ -319,88 +384,63 @@ static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
        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]);
+                               rs->unified_rw_rep ? "mixed" : 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)) {
-               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("    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);
+       if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
+               display_lat("slat", min, max, mean, dev);
+       if (calc_lat(&ts->clat_stat[ddir], &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);
 
-               free(minp);
-               free(maxp);
-       }
        if (ts->clat_percentiles) {
                show_clat_percentiles(ts->io_u_plat[ddir],
                                        ts->clat_stat[ddir].samples,
-                                       ts->percentile_list);
+                                       ts->percentile_list,
+                                       ts->percentile_precision);
        }
        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 (KB/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_lat(double *io_u_lat, int nr, const char **ranges,
-                    const char *msg)
+static int show_lat(double *io_u_lat, int nr, const char **ranges,
+                   const char *msg)
 {
-       int new_line = 1, i, line = 0;
+       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);
+                       log_info("    lat (%s) : ", msg);
                        new_line = 0;
                        line = 0;
                }
@@ -411,6 +451,11 @@ static void show_lat(double *io_u_lat, int nr, const char **ranges,
                if (line == 5)
                        new_line = 1;
        }
+
+       if (shown)
+               log_info("\n");
+
+       return shown;
 }
 
 static void show_lat_u(double *io_u_lat_u)
@@ -430,12 +475,16 @@ static void show_lat_m(double *io_u_lat_m)
        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)
+static void show_latencies(struct thread_stat *ts)
 {
+       double io_u_lat_u[FIO_IO_U_LAT_U_NR];
+       double io_u_lat_m[FIO_IO_U_LAT_M_NR];
+
+       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);
-       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)
@@ -443,30 +492,39 @@ 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_u[FIO_IO_U_LAT_U_NR];
-       double io_u_lat_m[FIO_IO_U_LAT_M_NR];
+       time_t time_p;
+       char time_buf[64];
 
-       if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
-           !(ts->total_io_u[0] + ts->total_io_u[1]))
+       if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
+           ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
+           ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
                return;
 
+       time(&time_p);
+       os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
+
        if (!ts->error) {
-               log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
+               log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
                                        ts->name, ts->groupid, ts->members,
-                                       ts->error, (int) ts->pid);
+                                       ts->error, (int) ts->pid, time_buf);
        } else {
-               log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
+               log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
                                        ts->name, ts->groupid, ts->members,
-                                       ts->error, ts->verror, (int) ts->pid);
+                                       ts->error, ts->verror, (int) ts->pid,
+                                       time_buf);
        }
 
-       if (ts->description)
+       if (strlen(ts->description))
                log_info("  Description  : [%s]\n", ts->description);
 
        if (ts->io_bytes[DDIR_READ])
                show_ddir_status(rs, ts, DDIR_READ);
        if (ts->io_bytes[DDIR_WRITE])
                show_ddir_status(rs, ts, DDIR_WRITE);
+       if (ts->io_bytes[DDIR_TRIM])
+               show_ddir_status(rs, ts, DDIR_TRIM);
+
+       show_latencies(ts);
 
        runtime = ts->total_run_time;
        if (runtime) {
@@ -482,7 +540,7 @@ void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
        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);
+       stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), 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],
@@ -501,14 +559,12 @@ void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
                                        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",
+       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]);
-       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,
@@ -521,16 +577,23 @@ 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;%llu", ts->io_bytes[ddir] >> 10, bw,
+               bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
+               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))
@@ -543,24 +606,153 @@ 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(";%f%%=%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);
 }
 
-#define FIO_TERSE_VERSION      "2"
+static void add_ddir_status_json(struct thread_stat *ts,
+               struct group_run_stats *rs, int ddir, struct json_object *parent)
+{
+       unsigned long min, max;
+       unsigned long long bw, iops;
+       unsigned int *ovals = NULL;
+       double mean, dev;
+       unsigned int len, minv, maxv;
+       int i;
+       const char *ddirname[] = {"read", "write", "trim"};
+       struct json_object *dir_object, *tmp_object, *percentile_object;
+       char buf[120];
+       double p_of_agg = 100.0;
 
-static void show_thread_status_terse(struct thread_stat *ts,
-                                    struct group_run_stats *rs)
+       assert(ddir_rw(ddir));
+
+       if (ts->unified_rw_rep && ddir != DDIR_READ)
+               return;
+
+       dir_object = json_create_object();
+       json_object_add_value_object(parent,
+               ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
+
+       iops = bw = 0;
+       if (ts->runtime[ddir]) {
+               uint64_t runt = ts->runtime[ddir];
+
+               bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
+               iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
+       }
+
+       json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
+       json_object_add_value_int(dir_object, "bw", bw);
+       json_object_add_value_int(dir_object, "iops", iops);
+       json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
+
+       if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
+               min = max = 0;
+               mean = dev = 0.0;
+       }
+       tmp_object = json_create_object();
+       json_object_add_value_object(dir_object, "slat", tmp_object);
+       json_object_add_value_int(tmp_object, "min", min);
+       json_object_add_value_int(tmp_object, "max", max);
+       json_object_add_value_float(tmp_object, "mean", mean);
+       json_object_add_value_float(tmp_object, "stddev", dev);
+
+       if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
+               min = max = 0;
+               mean = dev = 0.0;
+       }
+       tmp_object = json_create_object();
+       json_object_add_value_object(dir_object, "clat", tmp_object);
+       json_object_add_value_int(tmp_object, "min", min);
+       json_object_add_value_int(tmp_object, "max", max);
+       json_object_add_value_float(tmp_object, "mean", mean);
+       json_object_add_value_float(tmp_object, "stddev", dev);
+
+       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;
+
+       percentile_object = json_create_object();
+       json_object_add_value_object(tmp_object, "percentile", percentile_object);
+       for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
+               if (i >= len) {
+                       json_object_add_value_int(percentile_object, "0.00", 0);
+                       continue;
+               }
+               snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
+               json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
+       }
+
+       if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
+               min = max = 0;
+               mean = dev = 0.0;
+       }
+       tmp_object = json_create_object();
+       json_object_add_value_object(dir_object, "lat", tmp_object);
+       json_object_add_value_int(tmp_object, "min", min);
+       json_object_add_value_int(tmp_object, "max", max);
+       json_object_add_value_float(tmp_object, "mean", mean);
+       json_object_add_value_float(tmp_object, "stddev", dev);
+       if (ovals)
+               free(ovals);
+
+       if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
+               if (rs->agg[ddir]) {
+                       p_of_agg = mean * 100 / (double) rs->agg[ddir];
+                       if (p_of_agg > 100.0)
+                               p_of_agg = 100.0;
+               }
+       } else {
+               min = max = 0;
+               p_of_agg = mean = dev = 0.0;
+       }
+       json_object_add_value_int(dir_object, "bw_min", min);
+       json_object_add_value_int(dir_object, "bw_max", max);
+       json_object_add_value_float(dir_object, "bw_agg", mean);
+       json_object_add_value_float(dir_object, "bw_mean", mean);
+       json_object_add_value_float(dir_object, "bw_dev", dev);
+}
+
+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_u[FIO_IO_U_LAT_U_NR];
@@ -569,12 +761,13 @@ static void show_thread_status_terse(struct thread_stat *ts,
        int i;
 
        /* General Info */
-       log_info("%s;%s;%d;%d", FIO_TERSE_VERSION, ts->name, ts->groupid,
-                               ts->error);
+       log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
        /* Log Read Status */
-       show_ddir_status_terse(ts, rs, 0);
+       show_ddir_status_terse(ts, rs, DDIR_READ);
        /* Log Write Status */
-       show_ddir_status_terse(ts, rs, 1);
+       show_ddir_status_terse(ts, rs, DDIR_WRITE);
+       /* Log Trim Status */
+       show_ddir_status_terse(ts, rs, DDIR_TRIM);
 
        /* CPU Usage */
        if (ts->total_run_time) {
@@ -591,7 +784,7 @@ static void show_thread_status_terse(struct thread_stat *ts,
                                                                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_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
        stat_calc_lat_u(ts, io_u_lat_u);
        stat_calc_lat_m(ts, io_u_lat_m);
 
@@ -618,6 +811,166 @@ static void show_thread_status_terse(struct thread_stat *ts,
        log_info("\n");
 }
 
+static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
+                                          struct group_run_stats *rs, int ver)
+{
+       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("%d;%s;%s;%d;%d", ver, fio_version_string,
+                                       ts->name, ts->groupid, ts->error);
+       /* Log Read Status */
+       show_ddir_status_terse(ts, rs, DDIR_READ);
+       /* Log Write Status */
+       show_ddir_status_terse(ts, rs, DDIR_WRITE);
+       /* Log Trim Status */
+       if (ver == 4)
+               show_ddir_status_terse(ts, rs, DDIR_TRIM);
+
+       /* 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, ddir_rw_sum(ts->total_io_u), 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]);
+
+       /* disk util stats, if any */
+       show_disk_util(1, NULL);
+
+       /* 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);
+
+       /* Additional output if description is set */
+       if (strlen(ts->description))
+               log_info(";%s", ts->description);
+
+       log_info("\n");
+}
+
+static struct json_object *show_thread_status_json(struct thread_stat *ts,
+                                   struct group_run_stats *rs)
+{
+       struct json_object *root, *tmp;
+       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;
+
+       root = json_create_object();
+       json_object_add_value_string(root, "jobname", ts->name);
+       json_object_add_value_int(root, "groupid", ts->groupid);
+       json_object_add_value_int(root, "error", ts->error);
+
+       add_ddir_status_json(ts, rs, DDIR_READ, root);
+       add_ddir_status_json(ts, rs, DDIR_WRITE, root);
+       add_ddir_status_json(ts, rs, DDIR_TRIM, root);
+
+       /* 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;
+       }
+       json_object_add_value_float(root, "usr_cpu", usr_cpu);
+       json_object_add_value_float(root, "sys_cpu", sys_cpu);
+       json_object_add_value_int(root, "ctx", ts->ctx);
+       json_object_add_value_int(root, "majf", ts->majf);
+       json_object_add_value_int(root, "minf", ts->minf);
+
+
+       /* Calc % distribution of IO depths, usecond, msecond latency */
+       stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
+       stat_calc_lat_u(ts, io_u_lat_u);
+       stat_calc_lat_m(ts, io_u_lat_m);
+
+       tmp = json_create_object();
+       json_object_add_value_object(root, "iodepth_level", tmp);
+       /* Only show fixed 7 I/O depth levels*/
+       for (i = 0; i < 7; i++) {
+               char name[20];
+               if (i < 6)
+                       snprintf(name, 20, "%d", 1 << i);
+               else
+                       snprintf(name, 20, ">=%d", 1 << i);
+               json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
+       }
+
+       tmp = json_create_object();
+       json_object_add_value_object(root, "latency_us", tmp);
+       /* Microsecond latency */
+       for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
+               const char *ranges[] = { "2", "4", "10", "20", "50", "100",
+                                "250", "500", "750", "1000", };
+               json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
+       }
+       /* Millisecond latency */
+       tmp = json_create_object();
+       json_object_add_value_object(root, "latency_ms", tmp);
+       for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
+               const char *ranges[] = { "2", "4", "10", "20", "50", "100",
+                                "250", "500", "750", "1000", "2000",
+                                ">=2000", };
+               json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
+       }
+
+       /* Additional output if continue_on_error set - default off*/
+       if (ts->continue_on_error) {
+               json_object_add_value_int(root, "total_err", ts->total_err_count);
+               json_object_add_value_int(root, "total_err", ts->first_error);
+       }
+
+       /* Additional output if description is set */
+       if (strlen(ts->description))
+               json_object_add_value_string(root, "desc", ts->description);
+
+       return root;
+}
+
+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 || terse_version == 4)
+               show_thread_status_terse_v3_v4(ts, rs, terse_version);
+       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;
@@ -657,7 +1010,7 @@ void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
 {
        int i;
 
-       for (i = 0; i < 2; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; 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])
@@ -677,16 +1030,28 @@ 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);
+       for (l = 0; l < DDIR_RWDIR_CNT; l++) {
+               if (!dst->unified_rw_rep) {
+                       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];
+                       dst->io_bytes[l] += src->io_bytes[l];
 
-               if (dst->runtime[l] < src->runtime[l])
-                       dst->runtime[l] = src->runtime[l];
+                       if (dst->runtime[l] < src->runtime[l])
+                               dst->runtime[l] = src->runtime[l];
+               } else {
+                       sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
+                       sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
+                       sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
+                       sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
+
+                       dst->io_bytes[0] += src->io_bytes[l];
+
+                       if (dst->runtime[0] < src->runtime[l])
+                               dst->runtime[0] = src->runtime[l];
+               }
        }
 
        dst->usr_time += src->usr_time;
@@ -706,15 +1071,25 @@ void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
        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_RWDIR_CNT; k++) {
+               if (!dst->unified_rw_rep) {
+                       dst->total_io_u[k] += src->total_io_u[k];
+                       dst->short_io_u[k] += src->short_io_u[k];
+               } else {
+                       dst->total_io_u[0] += src->total_io_u[k];
+                       dst->short_io_u[0] += src->short_io_u[k];
+               }
        }
 
-       for (k = 0; k <= DDIR_WRITE; k++) {
+       for (k = 0; k < DDIR_RWDIR_CNT; 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];
+
+               for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
+                       if (!dst->unified_rw_rep)
+                               dst->io_u_plat[k][m] += src->io_u_plat[k][m];
+                       else
+                               dst->io_u_plat[0][m] += src->io_u_plat[k][m];
+               }
        }
 
        dst->total_run_time += src->total_run_time;
@@ -724,9 +1099,11 @@ void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
 
 void init_group_run_stat(struct group_run_stats *gs)
 {
+       int i;
        memset(gs, 0, sizeof(*gs));
-       gs->min_bw[0] = gs->min_run[0] = ~0UL;
-       gs->min_bw[1] = gs->min_run[1] = ~0UL;
+
+       for (i = 0; i < DDIR_RWDIR_CNT; i++)
+               gs->min_bw[i] = gs->min_run[i] = ~0UL;
 }
 
 void init_thread_stat(struct thread_stat *ts)
@@ -735,7 +1112,7 @@ void init_thread_stat(struct thread_stat *ts)
 
        memset(ts, 0, sizeof(*ts));
 
-       for (j = 0; j <= DDIR_WRITE; j++) {
+       for (j = 0; j < DDIR_RWDIR_CNT; j++) {
                ts->lat_stat[j].min_val = -1UL;
                ts->clat_stat[j].min_val = -1UL;
                ts->slat_stat[j].min_val = -1UL;
@@ -751,6 +1128,8 @@ void show_run_stats(void)
        struct thread_stat *threadstats, *ts;
        int i, j, nr_ts, last_ts, idx;
        int kb_base_warned = 0;
+       struct json_object *root = NULL;
+       struct json_array *array = NULL;
 
        runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
 
@@ -795,10 +1174,8 @@ 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));
+               ts->percentile_precision = td->o.percentile_precision;
+               memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
 
                idx++;
                ts->members++;
@@ -814,6 +1191,11 @@ void show_run_stats(void)
                        else
                                memset(ts->description, 0, FIO_JOBNAME_SIZE);
 
+                       /*
+                        * If multiple entries in this group, this is
+                        * the first member.
+                        */
+                       ts->thread_number = td->thread_number;
                        ts->groupid = td->groupid;
 
                        /*
@@ -822,6 +1204,7 @@ void show_run_stats(void)
                        ts->pid = td->pid;
 
                        ts->kb_base = td->o.kb_base;
+                       ts->unified_rw_rep = td->o.unified_rw_rep;
                } 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);
@@ -851,8 +1234,9 @@ void show_run_stats(void)
                ts = &threadstats[i];
                rs = &runstats[ts->groupid];
                rs->kb_base = ts->kb_base;
+               rs->unified_rw_rep += ts->unified_rw_rep;
 
-               for (j = 0; j <= DDIR_WRITE; j++) {
+               for (j = 0; j < DDIR_RWDIR_CNT; j++) {
                        if (!ts->runtime[j])
                                continue;
                        if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
@@ -862,10 +1246,11 @@ void show_run_stats(void)
 
                        bw = 0;
                        if (ts->runtime[j]) {
-                               unsigned long runt;
+                               unsigned long runt = ts->runtime[j];
+                               unsigned long long kb;
 
-                               runt = ts->runtime[j];
-                               bw = ts->io_bytes[j] / runt;
+                               kb = ts->io_bytes[j] / rs->kb_base;
+                               bw = kb * 1000 / runt;
                        }
                        if (bw < rs->min_bw[j])
                                rs->min_bw[j] = bw;
@@ -877,23 +1262,28 @@ void show_run_stats(void)
        }
 
        for (i = 0; i < groupid + 1; i++) {
-               unsigned long max_run[2];
+               int ddir;
 
                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] * 1000) / max_run[0];
-               if (rs->max_run[1])
-                       rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
+               for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
+                       if (rs->max_run[ddir])
+                               rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
+                                               rs->max_run[ddir];
+               }
        }
 
        /*
         * don't overwrite last signal output
         */
-       if (!terse_output)
+       if (output_format == FIO_OUTPUT_NORMAL)
                log_info("\n");
+       else if (output_format == FIO_OUTPUT_JSON) {
+               root = json_create_object();
+               json_object_add_value_string(root, "fio version", fio_version_string);
+               array = json_create_array();
+               json_object_add_value_array(root, "jobs", array);
+       }
 
        for (i = 0; i < nr_ts; i++) {
                ts = &threadstats[i];
@@ -901,30 +1291,108 @@ void show_run_stats(void)
 
                if (is_backend)
                        fio_server_send_ts(ts, rs);
-               else if (terse_output)
+               else if (output_format == FIO_OUTPUT_TERSE)
                        show_thread_status_terse(ts, rs);
-               else
+               else if (output_format == FIO_OUTPUT_JSON) {
+                       struct json_object *tmp = show_thread_status_json(ts, rs);
+                       json_array_add_value_object(array, tmp);
+               } else
                        show_thread_status(ts, rs);
        }
+       if (output_format == FIO_OUTPUT_JSON) {
+               /* disk util stats, if any */
+               show_disk_util(1, root);
 
-       if (!terse_output) {
-               for (i = 0; i < groupid + 1; i++) {
-                       rs = &runstats[i];
+               show_idle_prof_stats(FIO_OUTPUT_JSON, root);
 
-                       rs->groupid = i;
-                       if (is_backend)
-                               fio_server_send_gs(rs);
-                       else
-                               show_group_stats(rs);
-               }
+               json_print_object(root);
+               log_info("\n");
+               json_free_object(root);
+       }
+
+       for (i = 0; i < groupid + 1; i++) {
+               rs = &runstats[i];
+
+               rs->groupid = i;
+               if (is_backend)
+                       fio_server_send_gs(rs);
+               else if (output_format == FIO_OUTPUT_NORMAL)
+                       show_group_stats(rs);
+       }
 
-               show_disk_util();
+       if (is_backend)
+               fio_server_send_du();
+       else if (output_format == FIO_OUTPUT_NORMAL) {
+               show_disk_util(0, NULL);
+               show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
        }
 
        free(runstats);
        free(threadstats);
 }
 
+static void *__show_running_run_stats(void *arg)
+{
+       struct thread_data *td;
+       unsigned long long *rt;
+       struct timeval tv;
+       int i;
+
+       rt = malloc(thread_number * sizeof(unsigned long long));
+       fio_gettime(&tv, NULL);
+
+       for_each_td(td, i) {
+               rt[i] = mtime_since(&td->start, &tv);
+               if (td_read(td) && td->io_bytes[DDIR_READ])
+                       td->ts.runtime[DDIR_READ] += rt[i];
+               if (td_write(td) && td->io_bytes[DDIR_WRITE])
+                       td->ts.runtime[DDIR_WRITE] += rt[i];
+               if (td_trim(td) && td->io_bytes[DDIR_TRIM])
+                       td->ts.runtime[DDIR_TRIM] += rt[i];
+
+               td->update_rusage = 1;
+               td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
+               td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
+               td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
+               td->ts.total_run_time = mtime_since(&td->epoch, &tv);
+       }
+
+       for_each_td(td, i) {
+               if (td->rusage_sem) {
+                       td->update_rusage = 1;
+                       fio_mutex_down(td->rusage_sem);
+               }
+               td->update_rusage = 0;
+       }
+
+       show_run_stats();
+
+       for_each_td(td, i) {
+               if (td_read(td) && td->io_bytes[DDIR_READ])
+                       td->ts.runtime[DDIR_READ] -= rt[i];
+               if (td_write(td) && td->io_bytes[DDIR_WRITE])
+                       td->ts.runtime[DDIR_WRITE] -= rt[i];
+               if (td_trim(td) && td->io_bytes[DDIR_TRIM])
+                       td->ts.runtime[DDIR_TRIM] -= rt[i];
+       }
+
+       free(rt);
+       return NULL;
+}
+
+/*
+ * Called from signal handler. It _should_ be safe to just run this inline
+ * in the sig handler, but we should be disturbing the system less by just
+ * creating a thread to do it.
+ */
+void show_running_run_stats(void)
+{
+       pthread_t thread;
+
+       pthread_create(&thread, NULL, __show_running_run_stats, NULL);
+       pthread_detach(thread);
+}
+
 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
 {
        double val = data;
@@ -950,6 +1418,9 @@ static void __add_log_sample(struct io_log *iolog, unsigned long val,
 {
        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;
 
@@ -964,14 +1435,74 @@ static void __add_log_sample(struct io_log *iolog, unsigned long val,
        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 int bs)
 {
+       unsigned long elapsed, this_window;
+
        if (!ddir_rw(ddir))
                return;
 
-       __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch));
+       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);
+       }
+       if (iolog->avg_window[DDIR_TRIM].samples) {
+               unsigned long mw;
+
+               mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
+               __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed);
+       }
+
+
+       reset_io_stat(&iolog->avg_window[DDIR_READ]);
+       reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
+       reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
+       iolog->avg_last = elapsed;
 }
 
 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
@@ -1052,15 +1583,26 @@ void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
        if (spent < td->o.bw_avg_time)
                return;
 
-       rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) *
-                       1000 / spent / 1024;
-       add_stat_sample(&ts->bw_stat[ddir], rate);
+       /*
+        * Compute both read and write rates for the interval.
+        */
+       for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; 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 (td->bw_log)
+                       add_log_sample(td, td->bw_log, rate, ddir, bs);
 
-       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);
-       td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
 }
 
 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
@@ -1076,15 +1618,24 @@ void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
        if (spent < td->o.iops_avg_time)
                return;
 
-       iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent;
+       /*
+        * Compute both read and write rates for the interval.
+        */
+       for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; 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);
 
-       add_stat_sample(&ts->iops_stat[ddir], iops);
+               if (td->iops_log)
+                       add_log_sample(td, td->iops_log, iops, ddir, 0);
 
-       if (td->iops_log) {
-               assert(iops);
-               add_log_sample(td, td->iops_log, iops, ddir, 0);
+               td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
        }
 
        fio_gettime(&td->iops_sample_time, NULL);
-       td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
 }