* resend [PATCH] fix ramp_in
@ 2010-12-13 20:58 Steven Pratt
2010-12-14 7:33 ` Jens Axboe
0 siblings, 1 reply; 2+ messages in thread
From: Steven Pratt @ 2010-12-13 20:58 UTC (permalink / raw)
To: fio@vger.kernel.org, Jens Axboe
Let's try this again without the mangling.
There are a couple of problems with the relatively new ramp_in feature
of fio. First, the estimated time
to completion did not correctly take it into account and bounces around.
Second and more importantly,
the runtime was including ramp in time in throughput calculations even
though the IO done during that
time was ignored, thus making throughput metrics incorrect. This patch
fixes both.
Signed-off-by Steven Pratt <slpratt@austin.ibm.com>
eta.c | 4 ++--
fio.c | 12 +++++++-----
fio.h | 2 +-
stat.c | 4 ++--
4 files changed, 12 insertions(+), 10 deletions(-)
diff -Naur fio-1.44.1/eta.c fio-ramp/eta.c
--- fio-1.44.1/eta.c 2010-10-22 14:44:00.000000000 -0500
+++ fio-ramp/eta.c 2010-12-13 13:28:50.573808191 -0600
@@ -161,12 +161,12 @@
* if given, otherwise assume it'll run at the specified
rate.
*/
if (td->o.timeout)
{
- t_eta = td->o.timeout +
td->o.start_delay;
+ t_eta = td->o.timeout + td->o.start_delay
+td->o.ramp_time;
if (in_ramp_time(td))
{
unsigned long
ramp_left;
- ramp_left =
mtime_since_now(&td->start);
+ ramp_left =
mtime_since_now(&td->epoch);
ramp_left = (ramp_left + 999) /
1000;
if (ramp_left <=
t_eta)
t_eta -=
ramp_left;
diff -Naur fio-1.44.1/fio.c
fio-ramp/fio.c
--- fio-1.44.1/fio.c 2010-10-22 14:44:00.000000000
-0500
+++ fio-ramp/fio.c 2010-12-13 13:29:36.611258945
-0600
@@ -972,6 +972,8
@@
}
fio_gettime(&tv,
NULL);
+ td->ts.runtime[0] =
0;
+ td->ts.runtime[1] =
0;
memcpy(&td->epoch, &tv,
sizeof(tv));
memcpy(&td->start, &tv,
sizeof(tv));
}
@@ -1153,11 +1155,11
@@
if (td_read(td) && td->io_bytes[DDIR_READ])
{
elapsed =
utime_since_now(&td->start);
- runtime[DDIR_READ] +=
elapsed;
+ td->ts.runtime[DDIR_READ] +=
elapsed;
}
if (td_write(td) && td->io_bytes[DDIR_WRITE])
{
elapsed =
utime_since_now(&td->start);
- runtime[DDIR_WRITE] +=
elapsed;
+ td->ts.runtime[DDIR_WRITE] +=
elapsed;
}
if (td->error ||
td->terminate)
@@ -1174,15 +1176,15
@@
do_verify(td);
- runtime[DDIR_READ] +=
utime_since_now(&td->start);
+ td->ts.runtime[DDIR_READ] +=
utime_since_now(&td->start);
if (td->error ||
td->terminate)
break;
}
update_rusage_stat(td);
- td->ts.runtime[0] = (runtime[0] + 999) / 1000;
- td->ts.runtime[1] = (runtime[1] + 999) / 1000;
+ td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
+ td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
td->ts.total_run_time = mtime_since_now(&td->epoch);
td->ts.io_bytes[0] = td->io_bytes[0];
td->ts.io_bytes[1] = td->io_bytes[1];
diff -Naur fio-1.44.1/fio.h fio-ramp/fio.h
--- fio-1.44.1/fio.h 2010-10-22 14:44:00.000000000 -0500
+++ fio-ramp/fio.h 2010-10-29 16:30:55.000000000 -0500
@@ -126,7 +126,7 @@
unsigned long total_complete;
unsigned long long io_bytes[2];
- unsigned long runtime[2];
+ unsigned long long runtime[2];
unsigned long total_run_time;
/*
diff -Naur fio-1.44.1/stat.c fio-ramp/stat.c
--- fio-1.44.1/stat.c 2010-10-22 14:44:00.000000000 -0500
+++ fio-ramp/stat.c 2010-10-29 16:33:49.000000000 -0500
@@ -172,7 +172,7 @@
iops = (1000 * ts->total_io_u[ddir]) / runt;
iops_p = num2str(iops, 6, 1, 0);
- log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6lumsec\n",
+ 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]);
@@ -380,7 +380,7 @@
if (ts->runtime[ddir])
bw = ts->io_bytes[ddir] / ts->runtime[ddir];
- log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw,
+ log_info(";%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw,
ts->runtime[ddir]);
if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: resend [PATCH] fix ramp_in
2010-12-13 20:58 resend [PATCH] fix ramp_in Steven Pratt
@ 2010-12-14 7:33 ` Jens Axboe
0 siblings, 0 replies; 2+ messages in thread
From: Jens Axboe @ 2010-12-14 7:33 UTC (permalink / raw)
To: Steven Pratt; +Cc: fio@vger.kernel.org
On 2010-12-13 21:58, Steven Pratt wrote:
> Let's try this again without the mangling.
Actually this one is also badly mangled :-)
> There are a couple of problems with the relatively new ramp_in feature
> of fio. First, the estimated time
> to completion did not correctly take it into account and bounces around.
> Second and more importantly,
> the runtime was including ramp in time in throughput calculations even
> though the IO done during that
> time was ignored, thus making throughput metrics incorrect. This patch
> fixes both.
Thanks, I hand applied it. You could also remove runtime[] in
fio.c:thread_main() since it's now unused, I did that.
--
Jens Axboe
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2010-12-14 7:33 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2010-12-13 20:58 resend [PATCH] fix ramp_in Steven Pratt
2010-12-14 7:33 ` Jens Axboe
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox