Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 20 May 2013 11:49:13 -0600
From:      asomers@gmail.com
To:        florent+FreeBSD-hackers@peterschmitt.fr
Cc:        freebsd-hackers@freebsd.org
Subject:   Re: blogbench and write-open serialization
Message-ID:  <CAOtMX2jMj0cJwjFLq5BVdv-zKrkAaxY90T6uBQ8mDMmLMZe4xA@mail.gmail.com>
In-Reply-To: <51976D13.4010604@peterschmitt.fr>
References:  <kn6ujg$a46$1@ger.gmane.org> <51976D13.4010604@peterschmitt.fr>

next in thread | previous in thread | raw e-mail | index | archive | help
--001a11c2523e7fe4c604dd29f365
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable

On Sat, May 18, 2013 at 5:59 AM, Florent Peterschmitt
<florent@peterschmitt.fr> wrote:
> Le 18/05/2013 06:04, Ivan Voras a =E9crit :
>
>> During the BSDCan & DevSummit I got interested in finding out why
>> blogbench is so slow on FreeBSD. After talking to jhb, it looked like
>> one of the reasons might be that opening files with O_RDWR or O_WRONLY
>> (anything opening the file for writing) is serialized.
>>
>> To check this, I've written a small test program, which I've run on
>> CentOS 6.3 and FreeBSD 10-HEAD on the same hardware. Here are the result=
s:
>>
>> https://wiki.freebsd.org/Benchmarking/OpenCloseBenchmark
>>
>> Conclusions:
>>
>> * Linux opens and closes files much faster than FreeBSD
>> * Linux does not serialize write-open operations, while FreeBSD does
>> * Even with O_RDONLY, FreeBSD is much slower in opening (and closing)
>> files.
>>
>> I'd welcome a review of these results and comments.
>
>
> Hi,
>
> I'm no able to say anything about that (because I've no idea of how does
> Linux or FreeBSD works ), but could it be a problem from filesystem ?
>
> Everytime I had UFS I found the entire system very slow when doing some I=
/O
> (many little freezes), and with ZFS it's globally much better.
>
> _______________________________________________
> freebsd-hackers@freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
> To unsubscribe, send any mail to "freebsd-hackers-unsubscribe@freebsd.org=
"

I largely reproduced your results, but saw even worse scaling.
* Linux was much faster, for all thread counts and on all file systems
* Linux's performance degraded by about 25% at high thread counts.
FreeBSD's penalty was file-system dependent, but usually higher
* The filesystem and hard disk mattered little on linux (tested tmpfs
and ext4).  The hard disk mattered little on FreeBSD, but the file
system was very significant.
* On FreeBSD, tmpfs was the fasted for a single thread.  For nine
threads, tmpfs was fastest for writes (about 33% faster than UFS or
ZFS), but substantially slower than ZFS for reads.
* For O_RDONLY, ZFS scales much better than UFS as the number of
threads increases, but at low thread counts UFS was faster.

I was testing on a dual-socket, 6 core/socket Xeon E5645 @ 2.4GHz
machine.  For Linux, I used Debian 6.0.6 with the 2.6.32-5 kernel.
For FreeBSD, I was using a custom kernel based on stable/9.

--001a11c2523e7fe4c604dd29f365--



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAOtMX2jMj0cJwjFLq5BVdv-zKrkAaxY90T6uBQ8mDMmLMZe4xA>