Add total latency log
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 60d5260..0f259d2 100644 (file)
--- a/stat.c
+++ b/stat.c
 #include <string.h>
 #include <sys/time.h>
 #include <sys/types.h>
+#include <sys/stat.h>
 #include <dirent.h>
 #include <libgen.h>
 #include <math.h>
 
 #include "fio.h"
+#include "diskutil.h"
 
-static struct itimerval itimer;
-static LIST_HEAD(disk_list);
-
-static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
+void update_rusage_stat(struct thread_data *td)
 {
-       unsigned in_flight;
-       char line[256];
-       FILE *f;
-       char *p;
-
-       f = fopen(du->path, "r");
-       if (!f)
-               return 1;
-
-       p = fgets(line, sizeof(line), f);
-       if (!p) {
-               fclose(f);
-               return 1;
-       }
+       struct thread_stat *ts = &td->ts;
 
-       if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
-               fclose(f);
-               return 1;
-       }
+       getrusage(RUSAGE_SELF, &ts->ru_end);
 
-       fclose(f);
-       return 0;
+       ts->usr_time += mtime_since(&ts->ru_start.ru_utime,
+                                       &ts->ru_end.ru_utime);
+       ts->sys_time += mtime_since(&ts->ru_start.ru_stime,
+                                       &ts->ru_end.ru_stime);
+       ts->ctx += ts->ru_end.ru_nvcsw + ts->ru_end.ru_nivcsw
+                       - (ts->ru_start.ru_nvcsw + ts->ru_start.ru_nivcsw);
+       ts->minf += ts->ru_end.ru_minflt - ts->ru_start.ru_minflt;
+       ts->majf += ts->ru_end.ru_majflt - ts->ru_start.ru_majflt;
+
+       memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
 }
 
-static void update_io_tick_disk(struct disk_util *du)
+static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
+                   double *mean, double *dev)
 {
-       struct disk_util_stat __dus, *dus, *ldus;
-       struct timeval t;
+       double n = is->samples;
 
-       if (get_io_ticks(du, &__dus))
-               return;
+       if (is->samples == 0)
+               return 0;
 
-       dus = &du->dus;
-       ldus = &du->last_dus;
-
-       dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
-       dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
-       dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
-       dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
-       dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
-       dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
-       dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
-       dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
-       dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
-       dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
-
-       gettimeofday(&t, NULL);
-       du->msec += mtime_since(&du->time, &t);
-       memcpy(&du->time, &t, sizeof(t));
-       memcpy(ldus, &__dus, sizeof(__dus));
-}
+       *min = is->min_val;
+       *max = is->max_val;
 
-void update_io_ticks(void)
-{
-       struct list_head *entry;
-       struct disk_util *du;
+       n = (double) is->samples;
+       *mean = is->mean;
 
-       list_for_each(entry, &disk_list) {
-               du = list_entry(entry, struct disk_util, list);
-               update_io_tick_disk(du);
-       }
+       if (n > 1.0)
+               *dev = sqrt(is->S / (n - 1.0));
+       else
+               *dev = 0;
+
+       return 1;
 }
 
-static int disk_util_exists(dev_t dev)
+static void show_group_stats(struct group_run_stats *rs, int id)
 {
-       struct list_head *entry;
-       struct disk_util *du;
-
-       list_for_each(entry, &disk_list) {
-               du = list_entry(entry, struct disk_util, list);
+       char *p1, *p2, *p3, *p4;
+       const char *ddir_str[] = { "   READ", "  WRITE" };
+       int i;
 
-               if (du->dev == dev)
-                       return 1;
-       }
+       log_info("\nRun status group %d (all jobs):\n", id);
 
-       return 0;
-}
+       for (i = 0; i <= DDIR_WRITE; i++) {
+               const int i2p = is_power_of_2(rs->kb_base);
 
-static void disk_util_add(dev_t dev, char *path)
-{
-       struct disk_util *du = malloc(sizeof(*du));
+               if (!rs->max_run[i])
+                       continue;
 
-       memset(du, 0, sizeof(*du));
-       INIT_LIST_HEAD(&du->list);
-       sprintf(du->path, "%s/stat", path);
-       du->name = strdup(basename(path));
-       du->dev = dev;
+               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);
 
-       gettimeofday(&du->time, NULL);
-       get_io_ticks(du, &du->last_dus);
+               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]);
 
-       list_add_tail(&du->list, &disk_list);
+               free(p1);
+               free(p2);
+               free(p3);
+               free(p4);
+       }
 }
 
-static int check_dev_match(dev_t dev, char *path)
+#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)
 {
-       unsigned int major, minor;
-       char line[256], *p;
-       FILE *f;
-
-       f = fopen(path, "r");
-       if (!f) {
-               perror("open path");
-               return 1;
-       }
+       int i;
 
-       p = fgets(line, sizeof(line), f);
-       if (!p) {
-               fclose(f);
-               return 1;
+       /*
+        * Do depth distribution calculations
+        */
+       for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
+               if (total) {
+                       io_u_dist[i] = (double) map[i] / (double) total;
+                       io_u_dist[i] *= 100.0;
+                       if (io_u_dist[i] < 0.1 && map[i])
+                               io_u_dist[i] = 0.1;
+               } else
+                       io_u_dist[i] = 0.0;
        }
+}
+
+static void stat_calc_lat(struct thread_stat *ts, double *dst,
+                         unsigned int *src, int nr)
+{
+       unsigned long total = ts_total_io_u(ts);
+       int i;
 
-       if (sscanf(p, "%u:%u", &major, &minor) != 2) {
-               fclose(f);
-               return 1;
+       /*
+        * Do latency distribution calculations
+        */
+       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);
+}
 
-       if (((major << 8) | minor) == dev) {
-               fclose(f);
+static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean,
+                       double *dev)
+{
+       if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
+               *min /= 1000;
+               *max /= 1000;
+               *mean /= 1000.0;
+               *dev /= 1000.0;
                return 0;
        }
 
-       fclose(f);
        return 1;
 }
 
-static int find_block_dir(dev_t dev, char *path)
+static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
+                            int ddir)
 {
-       struct dirent *dir;
-       struct stat st;
-       int found = 0;
-       DIR *D;
+       const char *ddir_str[] = { "read ", "write" };
+       unsigned long min, max, runt;
+       unsigned long long bw, iops;
+       double mean, dev;
+       char *io_p, *bw_p, *iops_p;
+       int i2p;
 
-       D = opendir(path);
-       if (!D)
-               return 0;
+       if (!ts->runtime[ddir])
+               return;
 
-       while ((dir = readdir(D)) != NULL) {
-               char full_path[256];
+       i2p = is_power_of_2(rs->kb_base);
+       runt = ts->runtime[ddir];
 
-               if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
-                       continue;
-               if (!strcmp(dir->d_name, "device"))
-                       continue;
+       bw = (1000 * ts->io_bytes[ddir]) / runt;
+       io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
+       bw_p = num2str(bw, 6, 1, i2p);
 
-               sprintf(full_path, "%s/%s", path, dir->d_name);
+       iops = (1000 * ts->total_io_u[ddir]) / runt;
+       iops_p = num2str(iops, 6, 1, 0);
 
-               if (!strcmp(dir->d_name, "dev")) {
-                       if (!check_dev_match(dev, full_path)) {
-                               found = 1;
-                               break;
-                       }
-               }
+       log_info("  %s: io=%sB, bw=%sB/s, iops=%s, runt=%6lumsec\n",
+                                       ddir_str[ddir], io_p, bw_p, iops_p,
+                                       ts->runtime[ddir]);
 
-               if (stat(full_path, &st) == -1) {
-                       perror("stat");
-                       break;
-               }
+       free(io_p);
+       free(bw_p);
+       free(iops_p);
 
-               if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
-                       continue;
+       if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
+               const char *base = "(usec)";
+               char *minp, *maxp;
 
-               found = find_block_dir(dev, full_path);
-               if (found) {
-                       strcpy(path, full_path);
-                       break;
-               }
-       }
+               if (!usec_to_msec(&min, &max, &mean, &dev))
+                       base = "(msec)";
 
-       closedir(D);
-       return found;
-}
+               minp = num2str(min, 6, 1, 0);
+               maxp = num2str(max, 6, 1, 0);
 
-void init_disk_util(struct thread_data *td)
-{
-       struct stat st;
-       char foo[256], tmp[256];
-       dev_t dev;
-       char *p;
+               log_info("    slat %s: min=%s, max=%s, avg=%5.02f,"
+                        " stdev=%5.02f\n", base, minp, maxp, mean, dev);
 
-       if (!td->do_disk_util)
-               return;
+               free(minp);
+               free(maxp);
+       }
+       if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
+               const char *base = "(usec)";
+               char *minp, *maxp;
 
-       if (!stat(td->file_name, &st)) {
-               if (S_ISBLK(st.st_mode))
-                       dev = st.st_rdev;
-               else
-                       dev = st.st_dev;
-       } else {
-               /*
-                * must be a file, open "." in that path
-                */
-               strcpy(foo, td->file_name);
-               p = dirname(foo);
-               if (stat(p, &st)) {
-                       perror("disk util stat");
-                       return;
-               }
+               if (!usec_to_msec(&min, &max, &mean, &dev))
+                       base = "(msec)";
+
+               minp = num2str(min, 6, 1, 0);
+               maxp = num2str(max, 6, 1, 0);
 
-               dev = st.st_dev;
+               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 (disk_util_exists(dev))
-               return;
-               
-       sprintf(foo, "/sys/block");
-       if (!find_block_dir(dev, foo))
-               return;
+               if (!usec_to_msec(&min, &max, &mean, &dev))
+                       base = "(msec)";
 
-       /*
-        * If there's a ../queue/ directory there, we are inside a partition.
-        * Check if that is the case and jump back. For loop/md/dm etc we
-        * are already in the right spot.
-        */
-       sprintf(tmp, "%s/../queue", foo);
-       if (!stat(tmp, &st)) {
-               p = dirname(foo);
-               sprintf(tmp, "%s/queue", p);
-               if (stat(tmp, &st)) {
-                       fprintf(stderr, "unknown sysfs layout\n");
-                       return;
-               }
-               sprintf(foo, "%s", p);
+               minp = num2str(min, 6, 1, 0);
+               maxp = num2str(max, 6, 1, 0);
+
+               log_info("     lat %s: min=%s, max=%s, avg=%5.02f,"
+                        " stdev=%5.02f\n", base, minp, maxp, mean, dev);
+
+               free(minp);
+               free(maxp);
        }
+       if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
+               double p_of_agg;
 
-       td->sysfs_root = strdup(foo);
-       disk_util_add(dev, foo);
+               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);
+       }
 }
 
-void disk_util_timer_arm(void)
+static void show_lat(double *io_u_lat, int nr, const char **ranges,
+                    const char *msg)
 {
-       itimer.it_value.tv_sec = 0;
-       itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
-       setitimer(ITIMER_REAL, &itimer, NULL);
+       int new_line = 1, i, line = 0;
+
+       for (i = 0; i < nr; i++) {
+               if (io_u_lat[i] <= 0.0)
+                       continue;
+               if (new_line) {
+                       if (line)
+                               log_info("\n");
+                       log_info("     lat (%s): ", msg);
+                       new_line = 0;
+                       line = 0;
+               }
+               if (line)
+                       log_info(", ");
+               log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
+               line++;
+               if (line == 5)
+                       new_line = 1;
+       }
 }
 
-void update_rusage_stat(struct thread_data *td)
+static void show_lat_u(double *io_u_lat_u)
 {
-       if (!(td->runtime[0] + td->runtime[1]))
-               return;
-
-       getrusage(RUSAGE_SELF, &td->ru_end);
+       const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
+                                "250=", "500=", "750=", "1000=", };
 
-       td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
-       td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
-       td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
-
-       
-       memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
+       show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
 }
 
-static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
-                   double *mean, double *dev)
+static void show_lat_m(double *io_u_lat_m)
 {
-       double n;
+       const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
+                                "250=", "500=", "750=", "1000=", "2000=",
+                                ">=2000=", };
 
-       if (is->samples == 0)
-               return 0;
-
-       *min = is->min_val;
-       *max = is->max_val;
-
-       n = (double) is->samples;
-       *mean = (double) is->val / n;
-       *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
-       if (!(*min + *max) && !(*mean + *dev))
-               return 0;
-
-       return 1;
+       show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
 }
 
-static void show_group_stats(struct group_run_stats *rs, int id)
+static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
 {
-       printf("\nRun status group %d (all jobs):\n", id);
-
-       if (rs->max_run[DDIR_READ])
-               printf("   READ: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_kb[0] >> 10, rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
-       if (rs->max_run[DDIR_WRITE])
-               printf("  WRITE: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_kb[1] >> 10, rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
+       show_lat_u(io_u_lat_u);
+       log_info("\n");
+       show_lat_m(io_u_lat_m);
+       log_info("\n");
 }
 
-static void show_disk_util(void)
+static void show_thread_status(struct thread_stat *ts,
+                              struct group_run_stats *rs)
 {
-       struct disk_util_stat *dus;
-       struct list_head *entry;
-       struct disk_util *du;
-       double util;
+       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];
+
+       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, (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);
+       }
 
-       printf("\nDisk stats (read/write):\n");
+       if (ts->description)
+               log_info("  Description  : [%s]\n", ts->description);
 
-       list_for_each(entry, &disk_list) {
-               du = list_entry(entry, struct disk_util, list);
-               dus = &du->dus;
+       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);
 
-               util = (double) 100 * du->dus.io_ticks / (double) du->msec;
-               if (util > 100.0)
-                       util = 100.0;
+       runtime = ts->total_run_time;
+       if (runtime) {
+               double runt = (double) runtime;
 
-               printf("  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
+               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("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
+                " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
+
+       stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
+       log_info("  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
+                " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
+                                       io_u_dist[1], io_u_dist[2],
+                                       io_u_dist[3], io_u_dist[4],
+                                       io_u_dist[5], io_u_dist[6]);
+
+       stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
+       log_info("     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
+                " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
+                                       io_u_dist[1], io_u_dist[2],
+                                       io_u_dist[3], io_u_dist[4],
+                                       io_u_dist[5], io_u_dist[6]);
+       stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
+       log_info("     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
+                " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
+                                       io_u_dist[1], io_u_dist[2],
+                                       io_u_dist[3], io_u_dist[4],
+                                       io_u_dist[5], io_u_dist[6]);
+       log_info("     issued r/w: total=%lu/%lu, short=%lu/%lu\n",
+                                       ts->total_io_u[0], ts->total_io_u[1],
+                                       ts->short_io_u[0], ts->short_io_u[1]);
+       stat_calc_lat_u(ts, io_u_lat_u);
+       stat_calc_lat_m(ts, io_u_lat_m);
+       show_latencies(io_u_lat_u, io_u_lat_m);
+       if (ts->continue_on_error) {
+               log_info("     errors    : total=%lu, first_error=%d/<%s>\n",
+                                       ts->total_err_count,
+                                       ts->first_error,
+                                       strerror(ts->first_error));
        }
 }
 
-static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
-                            int ddir)
+static void show_ddir_status_terse(struct thread_stat *ts,
+                                  struct group_run_stats *rs, int ddir)
 {
-       char *ddir_str[] = { "read ", "write" };
        unsigned long min, max;
        unsigned long long bw;
        double mean, dev;
 
-       if (!td->runtime[ddir])
-               return;
+       bw = 0;
+       if (ts->runtime[ddir])
+               bw = ts->io_bytes[ddir] / ts->runtime[ddir];
+
+       log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw,
+                                                       ts->runtime[ddir]);
 
-       bw = td->io_bytes[ddir] / td->runtime[ddir];
-       printf("  %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
+       if (calc_lat(&ts->slat_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 (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
-               printf("    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
+       if (calc_lat(&ts->clat_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 (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
-               printf("    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
+       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 (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
+       if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
                double p_of_agg;
 
                p_of_agg = mean * 100 / (double) rs->agg[ddir];
-               printf("    bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
-       }
+               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(struct thread_data *td,
-                              struct group_run_stats *rs)
+
+static void show_thread_status_terse(struct thread_stat *ts,
+                                    struct group_run_stats *rs)
 {
+       double io_u_dist[FIO_IO_U_MAP_NR];
+       double io_u_lat_u[FIO_IO_U_LAT_U_NR];
+       double io_u_lat_m[FIO_IO_U_LAT_M_NR];
        double usr_cpu, sys_cpu;
+       int i;
 
-       if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
-               return;
-
-       printf("%s: (groupid=%d): err=%2d:\n",td->name, td->groupid, td->error);
+       log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
 
-       show_ddir_status(td, rs, td->ddir);
-       if (td->io_bytes[td->ddir ^ 1])
-               show_ddir_status(td, rs, td->ddir ^ 1);
+       show_ddir_status_terse(ts, rs, 0);
+       show_ddir_status_terse(ts, rs, 1);
 
-       if (td->runtime[0] + td->runtime[1]) {
-               double runt = td->runtime[0] + td->runtime[1];
+       if (ts->total_run_time) {
+               double runt = (double) ts->total_run_time;
 
-               usr_cpu = (double) td->usr_time * 100 / runt;
-               sys_cpu = (double) td->sys_time * 100 / runt;
+               usr_cpu = (double) ts->usr_time * 100 / runt;
+               sys_cpu = (double) ts->sys_time * 100 / runt;
        } else {
                usr_cpu = 0;
                sys_cpu = 0;
        }
 
-       printf("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
+       log_info(";%f%%;%f%%;%lu;%lu;%lu", 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_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]);
+
+       for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
+               log_info(";%3.2f%%", io_u_lat_u[i]);
+       for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
+               log_info(";%3.2f%%", io_u_lat_m[i]);
+       if (ts->continue_on_error)
+               log_info(";%lu;%d", ts->total_err_count, ts->first_error);
+       log_info("\n");
+
+       if (ts->description)
+               log_info(";%s", ts->description);
+
+       log_info("\n");
+}
+
+static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
+{
+       double mean, S;
+
+       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
+        */
+       if (nr == 1) {
+               mean = src->mean;
+               S = src->S;
+       } else {
+               mean = ((src->mean * (double) (nr - 1))
+                               + dst->mean) / ((double) nr);
+               S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
+       }
+
+       dst->mean = mean;
+       dst->S = S;
 }
 
 void show_run_stats(void)
 {
        struct group_run_stats *runstats, *rs;
        struct thread_data *td;
-       int i;
+       struct thread_stat *threadstats, *ts;
+       int i, j, k, l, nr_ts, last_ts, idx;
+       int kb_base_warned = 0;
 
        runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
 
@@ -386,88 +494,227 @@ void show_run_stats(void)
                rs->min_bw[1] = rs->min_run[1] = ~0UL;
        }
 
-       for (i = 0; i < thread_number; i++) {
-               unsigned long long rbw, wbw;
+       /*
+        * find out how many threads stats we need. if group reporting isn't
+        * enabled, it's one-per-td.
+        */
+       nr_ts = 0;
+       last_ts = -1;
+       for_each_td(td, i) {
+               if (!td->o.group_reporting) {
+                       nr_ts++;
+                       continue;
+               }
+               if (last_ts == td->groupid)
+                       continue;
+
+               last_ts = td->groupid;
+               nr_ts++;
+       }
 
-               td = &threads[i];
+       threadstats = malloc(nr_ts * sizeof(struct thread_stat));
 
-               if (td->error) {
-                       printf("%s: %s\n", td->name, td->verror);
-                       continue;
+       for (i = 0; i < nr_ts; i++) {
+               ts = &threadstats[i];
+
+               memset(ts, 0, sizeof(*ts));
+               for (j = 0; j <= DDIR_WRITE; j++) {
+                       ts->lat_stat[j].min_val = -1UL;
+                       ts->clat_stat[j].min_val = -1UL;
+                       ts->slat_stat[j].min_val = -1UL;
+                       ts->bw_stat[j].min_val = -1UL;
+               }
+               ts->groupid = -1;
+       }
+
+       j = 0;
+       last_ts = -1;
+       idx = 0;
+       for_each_td(td, i) {
+               if (idx && (!td->o.group_reporting ||
+                   (td->o.group_reporting && last_ts != td->groupid))) {
+                       idx = 0;
+                       j++;
                }
 
-               rs = &runstats[td->groupid];
-
-               if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
-                       rs->min_run[0] = td->runtime[0];
-               if (td->runtime[0] > rs->max_run[0])
-                       rs->max_run[0] = td->runtime[0];
-               if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
-                       rs->min_run[1] = td->runtime[1];
-               if (td->runtime[1] > rs->max_run[1])
-                       rs->max_run[1] = td->runtime[1];
-
-               rbw = wbw = 0;
-               if (td->runtime[0])
-                       rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
-               if (td->runtime[1])
-                       wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
-
-               if (rbw < rs->min_bw[0])
-                       rs->min_bw[0] = rbw;
-               if (wbw < rs->min_bw[1])
-                       rs->min_bw[1] = wbw;
-               if (rbw > rs->max_bw[0])
-                       rs->max_bw[0] = rbw;
-               if (wbw > rs->max_bw[1])
-                       rs->max_bw[1] = wbw;
-
-               rs->io_kb[0] += td->io_bytes[0] >> 10;
-               rs->io_kb[1] += td->io_bytes[1] >> 10;
+               last_ts = td->groupid;
+
+               ts = &threadstats[j];
+
+               idx++;
+               ts->members++;
+
+               if (ts->groupid == -1) {
+                       /*
+                        * These are per-group shared already
+                        */
+                       ts->name = td->o.name;
+                       ts->description = td->o.description;
+                       ts->groupid = td->groupid;
+
+                       /*
+                        * first pid in group, not very useful...
+                        */
+                       ts->pid = td->pid;
+
+                       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;
+               }
+
+               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;
+                               ts->verror = td->verror;
+                       } else  if (td->error) {
+                               ts->error = td->error;
+                               ts->verror = td->verror;
+                       }
+               }
+
+               for (l = 0; l <= DDIR_WRITE; l++) {
+                       sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
+                       sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
+                       sum_stat(&ts->lat_stat[l], &td->ts.lat_stat[l], idx);
+                       sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
+
+                       ts->stat_io_bytes[l] += td->ts.stat_io_bytes[l];
+                       ts->io_bytes[l] += td->ts.io_bytes[l];
+
+                       if (ts->runtime[l] < td->ts.runtime[l])
+                               ts->runtime[l] = td->ts.runtime[l];
+               }
+
+               ts->usr_time += td->ts.usr_time;
+               ts->sys_time += td->ts.sys_time;
+               ts->ctx += td->ts.ctx;
+               ts->majf += td->ts.majf;
+               ts->minf += td->ts.minf;
+
+               for (k = 0; k < FIO_IO_U_MAP_NR; k++)
+                       ts->io_u_map[k] += td->ts.io_u_map[k];
+               for (k = 0; k < FIO_IO_U_MAP_NR; k++)
+                       ts->io_u_submit[k] += td->ts.io_u_submit[k];
+               for (k = 0; k < FIO_IO_U_MAP_NR; k++)
+                       ts->io_u_complete[k] += td->ts.io_u_complete[k];
+               for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
+                       ts->io_u_lat_u[k] += td->ts.io_u_lat_u[k];
+               for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
+                       ts->io_u_lat_m[k] += td->ts.io_u_lat_m[k];
+
+
+               for (k = 0; k <= DDIR_WRITE; k++) {
+                       ts->total_io_u[k] += td->ts.total_io_u[k];
+                       ts->short_io_u[k] += td->ts.short_io_u[k];
+               }
+
+               ts->total_run_time += td->ts.total_run_time;
+               ts->total_submit += td->ts.total_submit;
+               ts->total_complete += td->ts.total_complete;
+       }
+
+       for (i = 0; i < nr_ts; i++) {
+               unsigned long long bw;
+
+               ts = &threadstats[i];
+               rs = &runstats[ts->groupid];
+               rs->kb_base = ts->kb_base;
+
+               for (j = 0; j <= DDIR_WRITE; j++) {
+                       if (!ts->runtime[j])
+                               continue;
+                       if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
+                               rs->min_run[j] = ts->runtime[j];
+                       if (ts->runtime[j] > rs->max_run[j])
+                               rs->max_run[j] = ts->runtime[j];
+
+                       bw = 0;
+                       if (ts->runtime[j]) {
+                               unsigned long runt;
+
+                               runt = ts->runtime[j];
+                               bw = ts->io_bytes[j] / runt;
+                       }
+                       if (bw < rs->min_bw[j])
+                               rs->min_bw[j] = bw;
+                       if (bw > rs->max_bw[j])
+                               rs->max_bw[j] = bw;
+
+                       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] * rs->kb_base) / 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] * rs->kb_base) / max_run[1];
        }
 
        /*
         * don't overwrite last signal output
         */
-       printf("\n");
+       if (!terse_output)
+               log_info("\n");
 
-       for (i = 0; i < thread_number; i++) {
-               td = &threads[i];
-               rs = &runstats[td->groupid];
+       for (i = 0; i < nr_ts; i++) {
+               ts = &threadstats[i];
+               rs = &runstats[ts->groupid];
 
-               show_thread_status(td, rs);
+               if (terse_output)
+                       show_thread_status_terse(ts, rs);
+               else
+                       show_thread_status(ts, rs);
        }
 
-       for (i = 0; i < groupid + 1; i++)
-               show_group_stats(&runstats[i], i);
+       if (!terse_output) {
+               for (i = 0; i < groupid + 1; i++)
+                       show_group_stats(&runstats[i], i);
+
+               show_disk_util();
+       }
 
-       show_disk_util();
+       free(runstats);
+       free(threadstats);
 }
 
-static inline void add_stat_sample(struct io_stat *is, unsigned long val)
+static inline void add_stat_sample(struct io_stat *is, unsigned long data)
 {
-       if (val > is->max_val)
-               is->max_val = val;
-       if (val < is->min_val)
-               is->min_val = val;
+       double val = data;
+       double delta;
+
+       if (data > is->max_val)
+               is->max_val = data;
+       if (data < is->min_val)
+               is->min_val = data;
+
+       delta = val - is->mean;
+       if (delta) {
+               is->mean += delta / (is->samples + 1.0);
+               is->S += delta * (val - is->mean);
+       }
 
-       is->val += val;
-       is->val_sq += val * val;
        is->samples++;
 }
 
-static void add_log_sample(struct thread_data *td, struct io_log *iolog,
-                          unsigned long val, int ddir)
+static void __add_log_sample(struct io_log *iolog, unsigned long val,
+                            enum fio_ddir ddir, unsigned int bs,
+                            unsigned long time)
 {
+       const int nr_samples = iolog->nr_samples;
+
        if (iolog->nr_samples == iolog->max_samples) {
                int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
 
@@ -475,44 +722,77 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
                iolog->max_samples <<= 1;
        }
 
-       iolog->log[iolog->nr_samples].val = val;
-       iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
-       iolog->log[iolog->nr_samples].ddir = ddir;
+       iolog->log[nr_samples].val = val;
+       iolog->log[nr_samples].time = time;
+       iolog->log[nr_samples].ddir = ddir;
+       iolog->log[nr_samples].bs = bs;
        iolog->nr_samples++;
 }
 
-void add_clat_sample(struct thread_data *td, int ddir, unsigned long msec)
+static void add_log_sample(struct thread_data *td, struct io_log *iolog,
+                          unsigned long val, enum fio_ddir ddir,
+                          unsigned int bs)
+{
+       __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch));
+}
+
+void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
+{
+       struct io_log *iolog = agg_io_log[ddir];
+
+       __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
+}
+
+void add_clat_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->clat_stat[ddir], usec);
+
+       if (ts->clat_log)
+               add_log_sample(td, ts->clat_log, usec, ddir, bs);
+}
+
+void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
+                    unsigned long usec, unsigned int bs)
 {
-       add_stat_sample(&td->clat_stat[ddir], msec);
+       struct thread_stat *ts = &td->ts;
+
+       add_stat_sample(&ts->slat_stat[ddir], usec);
 
-       if (td->clat_log)
-               add_log_sample(td, td->clat_log, msec, ddir);
+       if (ts->slat_log)
+               add_log_sample(td, ts->slat_log, usec, ddir, bs);
 }
 
-void add_slat_sample(struct thread_data *td, int ddir, unsigned long msec)
+void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
+                   unsigned long usec, unsigned int bs)
 {
-       add_stat_sample(&td->slat_stat[ddir], msec);
+       struct thread_stat *ts = &td->ts;
 
-       if (td->slat_log)
-               add_log_sample(td, td->slat_log, msec, ddir);
+       add_stat_sample(&ts->lat_stat[ddir], usec);
+
+       if (ts->lat_log)
+               add_log_sample(td, ts->lat_log, usec, ddir, bs);
 }
 
-void add_bw_sample(struct thread_data *td, int ddir)
+void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
+                  struct timeval *t)
 {
-       unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
+       struct thread_stat *ts = &td->ts;
+       unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
        unsigned long rate;
 
-       if (spent < td->bw_avg_time)
+       if (spent < td->o.bw_avg_time)
                return;
 
-       rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
-       add_stat_sample(&td->bw_stat[ddir], rate);
+       rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) *
+                       1000 / spent / 1024;
+       add_stat_sample(&ts->bw_stat[ddir], rate);
 
-       if (td->bw_log)
-               add_log_sample(td, td->bw_log, rate, ddir);
+       if (ts->bw_log)
+               add_log_sample(td, ts->bw_log, rate, ddir, bs);
 
-       gettimeofday(&td->stat_sample_time[ddir], NULL);
-       td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
+       fio_gettime(&ts->stat_sample_time[ddir], NULL);
+       ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
 }
-
-