Re: Best suiting OS

From: Karl Denninger <karl(at)denninger(dot)net>
To: Axel Rau <Axel(dot)Rau(at)chaos1(dot)de>
Cc: PostgreSQL Performance <pgsql-performance(at)postgresql(dot)org>
Subject: Re: Best suiting OS
Date: 2009-10-06 13:39:11
Message-ID: 4ACB487F.7070703@denninger.net
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-performance

Axel Rau wrote:
>
> Am 05.10.2009 um 23:44 schrieb Karl Denninger:
>
>> Turn on softupdates. Fsck is deferred and the system comes up almost
>> instantly even with TB-sized partitions; the fsck then cleans up the
>> cruft.
> Last time, I checked, there was a issue with background-fsck.
> I will give it a chance with my new 8.0 box.
> Do you have any experience with SSDs w/o BBUed Raidcontroller?
> Are they fast enough to ensure flash write out of drive cache at power
> failure after fsync ack?
>
> Axel
> ---
> axel(dot)rau(at)chaos1(dot)de PGP-Key:29E99DD6 +49 151 2300 9283 computing @
> chaos claudius
IMHO use the right tools for the job. In a DBMS environment where data
integrity is "the deal" this means a BBU'd RAID adapter.

SSDs have their own set of issues, at least at present..... For data
that is read-only (or nearly-so) and of size where it can fit on a SSD
they can provide VERY significant performance benefits, in that there is
no seek or latency delay. However, any write-significant application is
IMHO still better-suited to rotating media at present. This will change
I'm sure, but it is what it is as of this point in time.

I have yet to run into a problem with background-fsck on a
softupdate-set filesystem. In theory there is a potential issue with
drives that make their own decision on write-reordering; in practice on
a DBMS system you run with a BBU'd RAID controller and as such the
controller and system UPS should prevent this from being an issue.

One of the potential issues that needs to be kept in mind with any
critical application is that disks that have "intelligence" may choose
to re-order writes. This can bring trouble (data corruption) in any
application where a drive claims to have committed a block to stable
storage where in fact it only has it in its buffer RAM and has not
written it to a platter yet. The only reasonable solution to this
problem is to run backed-up power so as to mitigate the risk of power
disappearing at an inopportune time. Backed-up power brings other
advantages as well (as a quality UPS usually comes with significant
filtering and power conditioning) which refuses the up front risk of
failures and is thus IMHO mandatory for any system that carries data you
care about.

-- Karl

Attachment Content-Type Size
karl.vcf text/x-vcard 124 bytes

In response to

Responses

Browse pgsql-performance by date

  From Date Subject
Next Message Scott Marlowe 2009-10-06 14:26:13 Re: Speed / Server
Previous Message Joshua Tolley 2009-10-06 13:33:04 Re: Dumping + restoring a subset of a table?