Обсуждение: Re: filesystem option tuning

Hi!

On Mon, May 17, 2004 at 06:04:54PM +0100, Richard Huxton wrote:
> share-postgres@think42.com wrote:
> > [...]
>
> In no official capacity whatsoever, welcome aboard.

Thanks ;-)

> > There is just one point where I found the documentation lacking any
> > description and practical hints (as opposed to all other topics), namely
> > that of how to tune a setup for maximum performance regarding the layout of
> > partitions on hard-disks and their mount options.
>
> I'm not a Sun user, so I can't give any OS-specific notes, but in general:
>   - Don't bypass the filesystem, but feel free to tinker with mount
> options if you think it will help

Right, raw partitions are too low-level for me these days anyhow...
I assume that all postgres partitions can be mounted with noatime?

>   - If you can put WAL on separate disk(s), all the better.

Does that mean only the xlog, or also the clog? As far as I understand, the
clog contains some meta-information on the xlog, so presumably it is flushed
to disc synchronously together with the xlog? That would mean that they each
need a separate disk to prevent one disk having to seek too often...?

>   - Battery-backed write-cache for your SCSI controller can be a big
> performance win

I probably won't be able to get such a setup for this project; that's why I
am bothering about which disk will be seeking how often.

>   - Tablespaces _should_ be available in the next release of PG, we'll
> know for sure soon. That might make life simpler for you if you do want
> to spread your database around by hand,

Ok, I think tablespaces are not the important thing - at least for this
project of ours.

> > What I should add is that reliability is a premium for us, we do not want to
> > sacrifice integrity for speed, and that we are tuning for a high commit rate
> > of small, simple transactions...
>
> Make sure the WAL is on fast disks I'd suggest. At a guess that'll be
> your bottleneck.
>
> For more info, your best bet is to check the archives on the
> plpgsql-performance list, and then post there. People will probably want
> to know more about your database size/number of concurrent
> transactions/disk systems etc.

Here goes ... we are talking about a database cluster with two tables where
things are happening, one is a kind of log that is simply "appended" to and
will expect to reach a size of several million entries in the time window
that is kept, the other is a persistent backing of application data that
will mostly see read-modify-writes of single records. Two writers to the
history, one writer to the data table. The volume of data is not very high
and RAM is enough...

If any more information is required feel free to ask - I would really
appreciate getting this disk layout sorted out.

Thanks,
Colin

On Wednesday 19 May 2004 13:02, share-postgres@think42.com wrote:
> >   - If you can put WAL on separate disk(s), all the better.
>
> Does that mean only the xlog, or also the clog? As far as I understand, the
> clog contains some meta-information on the xlog, so presumably it is
> flushed to disc synchronously together with the xlog? That would mean that
> they each need a separate disk to prevent one disk having to seek too
> often...?

You can put clog and xlog on same drive. That should be enough in most cases.
xlog is written sequentially and never read back other than for recovery
after a crash. clog is typically 8KB or a page and should not be an IO
overhead even in high traffic databases.

> >   - Battery-backed write-cache for your SCSI controller can be a big
> > performance win
>
> I probably won't be able to get such a setup for this project; that's why I
> am bothering about which disk will be seeking how often.

As I said earlier, xlog is written sequentially and if I am not mistaken clog
as well. So there should not be much seeking if they are on a separate drive.

(Please correct me if I am wrong)

> >   - Tablespaces _should_ be available in the next release of PG, we'll
> > know for sure soon. That might make life simpler for you if you do want
> > to spread your database around by hand,
>
> Ok, I think tablespaces are not the important thing - at least for this
> project of ours.

Well, if you have tablespaces, you don't have to mess with symlinking
clog/xlog or use location facility which is bit rough. You should be able to
manage such a setup solely from postgresql. That is an advantage of
tablespaces.

> Here goes ... we are talking about a database cluster with two tables where
> things are happening, one is a kind of log that is simply "appended" to and
> will expect to reach a size of several million entries in the time window
> that is kept, the other is a persistent backing of application data that
> will mostly see read-modify-writes of single records. Two writers to the
> history, one writer to the data table. The volume of data is not very high
> and RAM is enough...

Even if you have enough RAM, you should use pg_autovacuum so that your tables
are in shape. This is especially required when your update/insert rate is
high.

If your history logs needs to be rotated, you can take advantage of the fact
that DDL's in postgresql are fully transacted. So you can drop the table in a
transaction but nobody will notice anything unless it is committed. Makes a
transparent rotation.

HTH

 Shridhar

Shridhar Daithankar <shridhar@frodo.hserus.net> writes:
> On Wednesday 19 May 2004 13:02, share-postgres@think42.com wrote:
> - If you can put WAL on separate disk(s), all the better.
>>
>> Does that mean only the xlog, or also the clog?

> You can put clog and xlog on same drive.

You can, but I think you shouldn't.  The entire argument for giving xlog
its own drive revolves around the fact that xlog is written
sequentially, and so if it has its own spindle then you have near-zero
seek requirements.  As soon as you give that drive any other work to do,
you start losing the low-seek property.

Now as Shridhar says, clog is not a very high-I/O-volume thing, so in
one sense it doesn't much matter which drive you put it on.  But it
seems to me that clog acts much more like ordinary table files than it
acts like xlog.

            regards, tom lane