Re: XFS write speed drop

[Jan Engelhardt <jengelh@linux01.gwdg.de>]

>> CASE 1: Copying from one disk to another
>> ========================================
>> Copying a compiled 2.6.17-rc4 tree; 306907 KB in 28566 files in 2090
>> directories.
>
>OK, we can call this a metadata intensive workload - lots of small
>files, lots of creates.  Barriers will hurt the most here, as we'd
>already have been log I/O bound most likely, and I'd expect barriers
>to only slow that further.
>
Yes and the most important thing is that someone made -o barrier the 
default and did not notice. Someone else? :-D

>Yep, note the user/sys shows no change, we're basically IO bound in
>both tests, and barriers are hurting (as expected).


>> CASE 2: Removing
>> ================
>> Remove the copies we created in case 1.
>> 
>> 15:45 (none):/tmp # mount /dev/hdc2 /D -o barrier
>> 15:45 (none):/D # time rm -Rf kernel-0
>> real	3m31.901s
>> user	0m0.050s
>> sys	0m3.140s
>
>versus:
>
>> 15:49 (none):/ # mount /dev/hdc2 /D -o nobarrier
>> 15:49 (none):/D # time rm -Rf kernel-1
>> 
>> real	0m53.471s
>> user	0m0.070s
>> sys	0m1.990s
>> 15:50 (none):/D # cd /
>> 15:50 (none):/ # umount /D
>
>Also metadata intensive, of course.  All the same issues as above,
>and the same techniques should be used to address it.
>
>Odd that the system time jumped here.  Roughly the same decrease in
>performance though (3-4x).

You may, or may not, take that serious. I only ran each test once (since 
3m31s vs 53s shows "enough" of the issue).

>So, I agree, you're seeing the cost of write barriers here, but I
>don't see anything unexpected (unfortunately for you, I guess).
>
>The FAQ entry above will explain why they're enabled by default.
>Try the v2 log change I suggested, hopefully that will mitigate
>the problem somewhat.  Alternatively, buy some decent hardware if
>you're after performance. ;)
>
Well as mentioned, -o nobarrier solved it, and that's it. I do not actually 
need barriers (or an UPS, to poke on another thread), because power 
failures are rather rare in Germany.


Jan Engelhardt
-- 
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/