Another reduction idea would be to open/read/write io logs v= ia gz frontend, defaulting to regular I/O where compression is either unava= ilable or undesired. Should be a big deal with large logs, which is often t= he case. I planned to take on this in my spare time couple of months back, = but it's still where it was at that time :(.

On Apr 2, 2014 8:30 PM, "Jens Axboe" &= lt;axboe@kernel.dk> wrote:

On 2014-04-02 14:38, Mark Nelson wrote:

On 04/02/2014 01:55 PM, Jens Axboe wrote:

On 04/02/2014 12:24 PM, Mark Nelson wrote:

On 04/02/2014 12:55 PM, Jens Axboe wrote:

On 04/02/2014 11:53 AM, Mark Nelson wrote:

On 04/02/2014 12:38 PM, Jens Axboe wrote:

On 04/02/2014 09:23 AM, Mark Nelson wrote:

I've been playing around with usng the various write_*_log
options in
fio to record run statistics. =C2=A0What I'm finding is that some of th= e
logs
have malformed lines in them. =C2=A0The following snippet is from the bw log:

811820, 12263, 0, 4194304
812619, 35929, 0, 4194304
818437, 8448, 0, 419430, 4194304
1256299, 160627, 0, 4194304
1256831, 154274, 0, 4194304

Notice that the malformed line appears to be a partial write that
gets
truncated. =C2=A0The fio command being used to run the tests is
autogenerated, so job files aren't used. =C2=A0here's the one that<= br> created
the above log snippet:

/home/nhm/bin/fio
--rw=3Dread
--ioengine=3Dlibaio
--runtime=3D7200
--ramp_time=3D0
--numjobs=3D1
--direct=3D1
--bs=3D4194304B
--iodepth=3D16
--size=3D9216M
--write_iops_log=3D/tmp/cbt/ceph/kvmrbdfio/osd_ra-00004096/cl= ient_ra-00000128/op_size-04194304/concurrent_procs-008/iodept= h-016/read/output

--write_bw_log=3D/tmp/cbt/ceph/kvmrbdfio/osd_ra-00004096/clie= nt_ra-00000128/op_size-04194304/concurrent_procs-008/iodepth-= 016/read/output

--write_lat_log=3D/tmp/cbt/ceph/kvmrbdfio/osd_ra-00004096/cli= ent_ra-00000128/op_size-04194304/concurrent_procs-008/iodepth= -016/read/output

--time_based
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-0
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-1
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-2
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-3
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-4
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-5
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-6
--name=3D/srv/rbdfio-`hostname -s`-0/cbt-kvmrbdfio-7

The --numjobs=3D1 line should be removed, but I believe it gets
overridden
anyway with multiple name parameters. =C2=A0I wonder if having multiple
name
parameters (ie multiple fio processes) is causing those processes to
all
try to write to the same log files, potentially at the same time?
Any
thoughts would be much appreicated!

What version of fio are you running?

f42557f (ie master from yesterday) with a slightly different fix for
the
basename issues that just got resolved.

I can verify now that running fio with a single --name flag appears to
generate logs without malformated lines.

OK, I'll take a look, We should already be serializing the writing to the same log files, but your example definitely looks like we are not.

BTW, the numjobs option applies to each --name section. Just to clear
that up from your original email. So you really get numjobs *
name_sections running.

Aha, thanks for clearing that up!

Is github the right place to create new issues? =C2=A0If so, want me to
create one for this?

Sure, you can go ahead and do that.

Actually, I'm going to wait after hearing your response below!

I'm puzzled by this report. The writeout of the logs is fully
serialized, there should be no interleaving. Looking at your output,
it's definitely either the output from two jobs, or mangled output from=
just one job. The corrupted line is at time 818437, and the next line is much further ahead. Normally you would expect to see:

500, some-bw, 0, blocksize =C2=A0 =C2=A0 =C2=A0 =C2=A0<- job X
1000, some-bw, 0, blocksize
[...]
end-time, some-bw, 0, blocksize =C2=A0 =C2=A0 =C2=A0 =C2=A0<- last entry= from job X
500, some-bw, 0, blocksize =C2=A0 =C2=A0 =C2=A0 =C2=A0<- job Y now start= s

Note that the log writing always appends. Did you possibly have this
half-done files in that location from a previous run?

I think this is possible. =C2=A0My tool is supposed to clean out old tmp directories but it was broken (and unnoticed because all other output
files were getting overwritten each time). =C2=A0Guess I should prioritize = a
long-standing plan to use UUIDs for all tests/directories to make this
kind of bug more obvious. :)

That makes me feel better, and does make a lot more sense! But since I was = looking at this code anyway, I could not help but be annoyed at how we seri= alize all the logs when writing them out. If the logs are big, it can take = a long time per log. So I added some code to only lock per file instead. Th= is means that if no log sharing is happening, we'll write all 5 logs in= parallel. A quick test here with a 5s null run which resulted in 4 logs (t= wo tiny, the bw and iops log) reduced the log writing from 26 seconds to 14= seconds. And that was just two 18M logs.

So even if we didn't have a real bug here, at least the logging is now = substantially faster.

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe fio" in=
the body of a message to majordomo@vger.kernel.org
More majordomo info at =C2=A0http://vger.kernel.org/majordomo-info.htm= l