Add unified_rw_reporting option
authorJens Axboe <axboe@kernel.dk>
Wed, 30 Jan 2013 11:56:23 +0000 (12:56 +0100)
committerJens Axboe <axboe@kernel.dk>
Wed, 30 Jan 2013 11:56:23 +0000 (12:56 +0100)
If this is set, then fio will sum and display just a single set
of statistics for any IO type. By default, fio accounts and reports
each data direction separately.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
HOWTO
client.c
eta.c
fio.1
fio.h
options.c
server.c
server.h
stat.c
stat.h

diff --git a/HOWTO b/HOWTO
index d5fa987..8073240 100644 (file)
--- a/HOWTO
+++ b/HOWTO
@@ -353,6 +353,12 @@ kb_base=int        The base unit for a kilobyte. The defacto base is 2^10, 1024.
                ten unit instead, for obvious reasons. Allow values are
                1024 or 1000, with 1024 being the default.
 
+unified_rw_reporting=bool      Fio normally reports statistics on a per
+               data direction basis, meaning that read, write, and trim are
+               accounted and reported separately. If this option is set,
+               the fio will sum the results and report them as "mixed"
+               instead.
+
 randrepeat=bool        For random IO workloads, seed the generator in a predictable
                way so that results are repeatable across repetitions.
 
index 6021e4a..0dc620d 100644 (file)
--- a/client.c
+++ b/client.c
@@ -595,6 +595,7 @@ static void convert_ts(struct thread_stat *dst, struct thread_stat *src)
        dst->groupid    = le32_to_cpu(src->groupid);
        dst->pid        = le32_to_cpu(src->pid);
        dst->members    = le32_to_cpu(src->members);
+       dst->unified_rw_rep     = le32_to_cpu(src->unified_rw_rep);
 
        for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                convert_io_stat(&dst->clat_stat[i], &src->clat_stat[i]);
@@ -667,6 +668,7 @@ static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
 
        dst->kb_base    = le32_to_cpu(src->kb_base);
        dst->groupid    = le32_to_cpu(src->groupid);
+       dst->unified_rw_rep     = le32_to_cpu(src->unified_rw_rep);
 }
 
 static void handle_ts(struct fio_client *client, struct fio_net_cmd *cmd)
@@ -687,6 +689,7 @@ static void handle_ts(struct fio_client *client, struct fio_net_cmd *cmd)
 
        client_ts.members++;
        client_ts.groupid = p->ts.groupid;
+       client_ts.unified_rw_rep = p->ts.unified_rw_rep;
 
        if (++sum_stat_nr == sum_stat_clients) {
                strcpy(client_ts.name, "All clients");
diff --git a/eta.c b/eta.c
index fdf55c5..39fe10f 100644 (file)
--- a/eta.c
+++ b/eta.c
@@ -226,7 +226,8 @@ static int thread_eta(struct thread_data *td)
        return eta_sec;
 }
 
-static void calc_rate(unsigned long mtime, unsigned long long *io_bytes,
+static void calc_rate(int unified_rw_rep, unsigned long mtime,
+                     unsigned long long *io_bytes,
                      unsigned long long *prev_io_bytes, unsigned int *rate)
 {
        int i;
@@ -235,19 +236,32 @@ static void calc_rate(unsigned long mtime, unsigned long long *io_bytes,
                unsigned long long diff;
 
                diff = io_bytes[i] - prev_io_bytes[i];
-               rate[i] = ((1000 * diff) / mtime) / 1024;
+               if (unified_rw_rep) {
+                       rate[i] = 0;
+                       rate[0] += ((1000 * diff) / mtime) / 1024;
+               } else
+                       rate[i] = ((1000 * diff) / mtime) / 1024;
 
                prev_io_bytes[i] = io_bytes[i];
        }
 }
 
-static void calc_iops(unsigned long mtime, unsigned long long *io_iops,
+static void calc_iops(int unified_rw_rep, unsigned long mtime,
+                     unsigned long long *io_iops,
                      unsigned long long *prev_io_iops, unsigned int *iops)
 {
        int i;
 
        for (i = 0; i < DDIR_RWDIR_CNT; i++) {
-               iops[i] = ((io_iops[i] - prev_io_iops[i]) * 1000) / mtime;
+               unsigned long long diff;
+
+               diff = io_iops[i] - prev_io_iops[i];
+               if (unified_rw_rep) {
+                       iops[i] = 0;
+                       iops[0] += (diff * 1000) / mtime;
+               } else
+                       iops[i] = (diff * 1000) / mtime;
+
                prev_io_iops[i] = io_iops[i];
        }
 }
@@ -259,7 +273,7 @@ static void calc_iops(unsigned long mtime, unsigned long long *io_iops,
 int calc_thread_status(struct jobs_eta *je, int force)
 {
        struct thread_data *td;
-       int i;
+       int i, unified_rw_rep;
        unsigned long rate_time, disp_time, bw_avg_time, *eta_secs;
        unsigned long long io_bytes[DDIR_RWDIR_CNT];
        unsigned long long io_iops[DDIR_RWDIR_CNT];
@@ -293,7 +307,9 @@ int calc_thread_status(struct jobs_eta *je, int force)
        io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
        io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
        bw_avg_time = ULONG_MAX;
+       unified_rw_rep = 0;
        for_each_td(td, i) {
+               unified_rw_rep += td->o.unified_rw_rep;
                if (is_power_of_2(td->o.kb_base))
                        je->is_pow2 = 1;
                if (td->o.bw_avg_time < bw_avg_time)
@@ -339,9 +355,15 @@ int calc_thread_status(struct jobs_eta *je, int force)
 
                if (td->runstate > TD_RAMP) {
                        int ddir;
+
                        for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
-                               io_bytes[ddir] += td->io_bytes[ddir];
-                               io_iops[ddir] += td->io_blocks[ddir];
+                               if (unified_rw_rep) {
+                                       io_bytes[0] += td->io_bytes[ddir];
+                                       io_iops[0] += td->io_blocks[ddir];
+                               } else {
+                                       io_bytes[ddir] += td->io_bytes[ddir];
+                                       io_iops[ddir] += td->io_blocks[ddir];
+                               }
                        }
                }
        }
@@ -367,7 +389,8 @@ int calc_thread_status(struct jobs_eta *je, int force)
        rate_time = mtime_since(&rate_prev_time, &now);
 
        if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
-               calc_rate(rate_time, io_bytes, rate_io_bytes, je->rate);
+               calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
+                               je->rate);
                memcpy(&rate_prev_time, &now, sizeof(now));
                add_agg_sample(je->rate[DDIR_READ], DDIR_READ, 0);
                add_agg_sample(je->rate[DDIR_WRITE], DDIR_WRITE, 0);
@@ -382,8 +405,8 @@ int calc_thread_status(struct jobs_eta *je, int force)
        if (!force && disp_time < 900)
                return 0;
 
-       calc_rate(disp_time, io_bytes, disp_io_bytes, je->rate);
-       calc_iops(disp_time, io_iops, disp_io_iops, je->iops);
+       calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
+       calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
 
        memcpy(&disp_prev_time, &now, sizeof(now));
 
diff --git a/fio.1 b/fio.1
index 2f7728a..1f81ea7 100644 (file)
--- a/fio.1
+++ b/fio.1
@@ -240,6 +240,11 @@ The base unit for a kilobyte. The defacto base is 2^10, 1024.  Storage
 manufacturers like to use 10^3 or 1000 as a base ten unit instead, for obvious
 reasons. Allow values are 1024 or 1000, with 1024 being the default.
 .TP
+.BI unified_rw_reporting \fR=\fPbool
+Fio normally reports statistics on a per data direction basis, meaning that
+read, write, and trim are accounted and reported separately. If this option is
+set, the fio will sum the results and report them as "mixed" instead.
+.TP
 .BI randrepeat \fR=\fPbool
 Seed the random number generator in a predictable way so results are repeatable
 across runs.  Default: true.
diff --git a/fio.h b/fio.h
index 9e20299..2fd354a 100644 (file)
--- a/fio.h
+++ b/fio.h
@@ -237,6 +237,7 @@ struct thread_options {
        unsigned int disable_clat;
        unsigned int disable_slat;
        unsigned int disable_bw;
+       unsigned int unified_rw_rep;
        unsigned int gtod_reduce;
        unsigned int gtod_cpu;
        unsigned int gtod_offload;
index 1009df3..799e77a 100644 (file)
--- a/options.c
+++ b/options.c
@@ -2496,6 +2496,13 @@ static struct fio_option options[FIO_MAX_OPTS] = {
                .help   = "Set up dedicated gettimeofday() thread on this CPU",
                .verify = gtod_cpu_verify,
        },
+       {
+               .name   = "unified_rw_reporting",
+               .type   = FIO_OPT_BOOL,
+               .off1   = td_var_offset(unified_rw_rep),
+               .help   = "Unify reporting across data direction",
+               .def    = "0",
+       },
        {
                .name   = "continue_on_error",
                .type   = FIO_OPT_STR,
index 0cc3fad..7ec8531 100644 (file)
--- a/server.c
+++ b/server.c
@@ -648,6 +648,7 @@ static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
 
        dst->kb_base    = cpu_to_le32(src->kb_base);
        dst->groupid    = cpu_to_le32(src->groupid);
+       dst->unified_rw_rep     = cpu_to_le32(src->unified_rw_rep);
 }
 
 /*
@@ -669,8 +670,10 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
 
        p.ts.error      = cpu_to_le32(ts->error);
        p.ts.groupid    = cpu_to_le32(ts->groupid);
+       p.ts.unified_rw_rep     = cpu_to_le32(ts->unified_rw_rep);
        p.ts.pid        = cpu_to_le32(ts->pid);
        p.ts.members    = cpu_to_le32(ts->members);
+       p.ts.unified_rw_rep     = cpu_to_le32(ts->unified_rw_rep);
 
        for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                convert_io_stat(&p.ts.clat_stat[i], &ts->clat_stat[i]);
index 3f1bde4..15b802b 100644 (file)
--- a/server.h
+++ b/server.h
@@ -36,7 +36,7 @@ struct fio_net_int_cmd {
 };
 
 enum {
-       FIO_SERVER_VER          = 9,
+       FIO_SERVER_VER          = 10,
 
        FIO_SERVER_MAX_PDU      = 1024,
 
diff --git a/stat.c b/stat.c
index 8e1034b..5cc6b7b 100644 (file)
--- a/stat.c
+++ b/stat.c
@@ -277,9 +277,9 @@ 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);
@@ -381,8 +381,8 @@ 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);
@@ -695,8 +695,12 @@ static void add_ddir_status_json(struct thread_stat *ts,
 
        assert(ddir_rw(ddir));
 
+       if (ts->unified_rw_rep && ddir != DDIR_READ)
+               return;
+
        dir_object = json_create_object();
-       json_object_add_value_object(parent, ddirname[ddir], dir_object);
+       json_object_add_value_object(parent,
+               ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 
        iops = bw = 0;
        if (ts->runtime[ddir]) {
@@ -1062,15 +1066,27 @@ void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
        int l, k;
 
        for (l = 0; l < DDIR_RWDIR_CNT; l++) {
-               sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
-               sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
-               sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
-               sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
-
-               dst->io_bytes[l] += src->io_bytes[l];
-
-               if (dst->runtime[l] < src->runtime[l])
-                       dst->runtime[l] = src->runtime[l];
+               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];
+
+                       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;
@@ -1091,14 +1107,24 @@ void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
                dst->io_u_lat_m[k] += src->io_u_lat_m[k];
 
        for (k = 0; k < DDIR_RWDIR_CNT; k++) {
-               dst->total_io_u[k] += src->total_io_u[k];
-               dst->short_io_u[k] += src->short_io_u[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_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;
@@ -1210,6 +1236,7 @@ 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);
@@ -1239,6 +1266,7 @@ 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_RWDIR_CNT; j++) {
                        if (!ts->runtime[j])
diff --git a/stat.h b/stat.h
index 4ca8261..97186c1 100644 (file)
--- a/stat.h
+++ b/stat.h
@@ -8,6 +8,7 @@ struct group_run_stats {
        uint64_t agg[DDIR_RWDIR_CNT];
        uint32_t kb_base;
        uint32_t groupid;
+       uint32_t unified_rw_rep;
 };
 
 /*
@@ -120,6 +121,7 @@ struct thread_stat {
        uint32_t pid;
        char description[FIO_JOBNAME_SIZE];
        uint32_t members;
+       uint32_t unified_rw_rep;
 
        /*
         * bandwidth and latency stats