Re: ufs multilabel performance (fwd)

On Apr 15, 2012, at 1:17 PM, O. Hartmann wrote:

> Am 04/15/12 22:00, schrieb Garrett Cooper:
>> On Apr 15, 2012, at 12:30 PM, O. Hartmann wrote:
>> 
>>> Am 04/15/12 15:59, schrieb Richard Kojedzinszky:
>>>> Thank you for the reply.
>>>> 
>>>> Unfortunately, dont know why, but on my xen virtualised environment,
>>>> fbsd amd64 domU performs much slower, not only 30 times. Without
>>>> multilabel, file creation speed is around 2500/s, but with multilabels
>>>> enabled, it is only 15/s (!). so it is more than 100 times slower.
>>>> 
>>>> And anyway freebsd is known to be fast as well, as functional. The power
>>>> to serve. :)
>>>> 
>>>> But in my environment, 15/s file creation is very-very slow. The
>>>> hardware is a q6700 cpu with 4G ram, 2x1T sata disks in raid1, the host
>>>> runs linux. I think with this hw the mentioned speed is really slow.
>>>> 
>>>> Regards,
>>>> 
>>>> 
>>>> Kojedzinszky Richard
>>>> Euronet Magyarorszag Informatikai Zrt.
>>>> 
>>>> On Sun, 15 Apr 2012, O. Hartmann wrote:
>>>> 
>>>>> Date: Sun, 15 Apr 2012 13:20:23 +0200
>>>>> From: O. Hartmann <ohartman_at_zedat.fu-berlin.de>
>>>>> To: Richard Kojedzinszky <krichy_at_tvnetwork.hu>
>>>>> Cc: freebsd-security_at_freebsd.org
>>>>> Subject: Re: ufs multilabel performance (fwd)
>>>>> 
>>>>> Am 04/14/12 21:37, schrieb Richard Kojedzinszky:
>>>>>> Dear list,
>>>>>> 
>>>>>> Although it is not only security-related question, I did not get any
>>>>>> answer from freebsd-performance. The original question is below.
>>>>>> 
>>>>>> Can someone give some advice?
>>>>>> 
>>>>>> Thanks in advance,
>>>>>> 
>>>>>> 
>>>>>> Kojedzinszky Richard
>>>>>> Euronet Magyarorszag Informatikai Zrt.
>>>>>> 
>>>>>> ---------- Forwarded message ----------
>>>>>> Date: Thu, 10 Nov 2011 06:16:57 +0100 (CET)
>>>>>> From: Richard Kojedzinszky <krichy_at_tvnetwork.hu>
>>>>>> To: freebsd-performance_at_freebsd.org
>>>>>> Subject: ufs multilabel performance
>>>>>> 
>>>>>> Dear List,
>>>>>> 
>>>>>> I've noticed that when I enable multilabel on an fs, a file creation
>>>>>> gets around 20-30 times slower than without multilabel set.
>>>>>> 
>>>>>> This one-liner can be used to test the differences:
>>>>>> $ truss -D perl -e 'open(F, ">$_.file") for 1 .. 1000'
>>>>> 
>>>>> Same here, creating files seems to be 10 - 30 times slower with
>>>>> multilabels as it is without.
>>>>> 
>>>>> But as several posts and discussions reflects, FreeBSD isn't supposed to
>>>>> be fast although it is claimed that writing is the major than reading;
>>>>> FBSD should serve functionality.
>>>>>> 
>>>>>> And one can see that the open call takes much more when multilabel is
>>>>>> set on an fs. It seems that only file creation needs that many time,
>>>>>> when a file exists it is opened much faster.
>>>>>> 
>>>>>> Could someone acknowledge this, and have some suggestions how to make it
>>>>>> faster?
>>>>>> 
>>>>>> Regards,
>>>>>> 
>>>>>> 
>>>>>> Kojedzinszky Richard
>>>>>> TvNetWork Nyrt.
>>>>>> E-mail: krichy (at) tvnetwork [dot] hu
>>>>>> PGP: 0x54B2BF0C8F59B1B7
>>>>>> Fingerprint = F6D4 3FFE AF03 CACF 0DCB  46A1 54B2 BF0C 8F59 B1B7
>>> 
>>> At the moment, I'm troubled with a nasty kernel bug on all FreeBSD 10
>>> boxes I have spare to test.
>>> 
>>> I just tried to reproduce your observation and as far as I can go with
>>> my experience, I can confirm that by using your perl script.
>>> 
>>> I'd like to test this again with a small C program.
>>> 
>>> I can only test the issue (test is too far optimistic, it's simply a
>>> reproduction of your observation) on FreeBSD 10, the only remaining
>>> FreeBSD server at our department is running FBSD 9-STABLE/amd64 and "in
>>> production", so changing multilabel support is a bit harsh at the moment.
>>> 
>>> 
>>> Sorry about crossposting, but I think this belongs more to CURRENT and
>>> PERFORMANCE than SECURITY.
>> 
>> My suggestion is completely take perl out of the equation because the way you're invoking it above uses stdio and a few other things that add unnecessary overhead.
>> 
>> Try the attached C program/bourne shell snippet instead.
>> 
>> Cheers,
>> -Garrett
>> 
>> #!/bin/sh
>> 
>> set -e
>> 
>> tmp=$(mktemp -d tmp.XXXXXX)
>> trap "cd /; rm -Rf $tmp" EXIT
>> cd $tmp
>> 
>> cat > test_open.c <<EOF
>> 
>> #include <fcntl.h>
>> #include <stdio.h>
>> #include <unistd.h>
>> 
>> int
>> main(void)
>> {
>>        char buf[20];
>>        int i;
>> 
>>        for (i = 0; i < 1000; i++) {
>>                sprintf(buf, "%d", i);
>>                close(open(buf, O_WRONLY|O_CREAT, 0600));
>>        }
>>        return (0);
>> }
>> EOF
>> 
>> gcc -o test_open test_open.c
>> time ./test_open_______________________________________________
> 
> This was pretty fast ;-)

	If it turns out that it wasn't that particular item that's causing a slowdown, I can easily modify my above snippet to use stdio, etc. But unless the version of perl and a few other items are the same, I wouldn't necessarily blame the guest OS. Please also make sure that Xen, etc is completely up-to-date because there were some performance degradation issues that were fixed post-6.0.
-Garrett