On 10/5/09 10:27 AM, "Karl Denninger" <k...@denninger.net> wrote:
> Scott Carey wrote:
>>
>>
>> On 10/3/09 7:35 PM, "Karl Denninger" <k...@denninger.net>
>> <mailto:k...@denninger.net> wrote:
>>
>>
>>
>>>
>>> I am a particular fan of FreeBSD, and in some benchmarking I did between it
>>> and CentOS FreeBSD 7.x literally wiped the floor with the CentOS release I
>>> tried on IDENTICAL hardware.
>>> I also like the 3ware raid coprocessors - they work well, are fast, and I've
>>> had zero trouble with them.
>>>
>>> -- Karl
>>>
>>>
>>
>>
>> With CentOS 5.x, I have to do quite a bit of tuning to get it to perform
>> well. I often get almost 2x the performance after tuning.
>>
>> For I/O --
>> Deadline scheduler + reasonably large block device read-ahead + XFS
>> configured with large 'allocsize' settings (8MB to 80MB) make a huge
>> difference.
>>
>> Furthermore, the 3ware 35xx and 36xx (I think) I tried performed
>> particularly badly out of the box without tuning on CentOS.
>>
>> So, Identical hardware or not, both have to be tuned well to really compare
>> anyway.
>>
>> However, I have certainly seen some inefficiencies with Linux and large use
>> of shared memory -- and I wouldn't be surprised if these problems don't
>> exist on FreeBSD or OpenSolaris.
>>
> I don't run the 3x series 3ware boards. If I recall correctly they're not
> true coprocessor boards and rely on the host CPU. Those are always going to
> be a lose compared to a true coprocessor with dedicated cache memory on the
> card.
I screwed up, it was the 95xx and 96xx that stink for me. (Adaptec 2x as
fast, PERC 6 25% faster) with 1TB SATA drives.
Thought 96xx was a good chunk faster due to the faster interface.
>
> The 9xxx series boards are, and are extremely fast (make sure you install the
> battery backup or run on a UPS, set the appropriate flags, and take your
> chances - writeback caching makes a HUGE difference.)
Not at all in my experience, 12 drives in raid 10, and 300MB/sec sequential
trasfer rate = crap. Heavily tweaked, 450MB/sec. (Adaptec 5805 =
600MB/sec).
>
> Other than pinning shared memory on FreeBSD (and increasing a couple of
> boot-time tunables to permit large enough shared segments and semaphore lists)
> little is required to get excellent performance.
>
> The LSI cards that DELL, Intel and a few others have used (these appear to be
> deprecated now as it looks like LSI bought 3ware) also work well but their
> user interface is somewhat of a pain in the butt compared to 3Ware's.
>
> -- Karl
>
--
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance