Обсуждение: Moving from 32 to 64 bit builds on Solaris

I've got an install of Postgres 8.2.3 on a Sun box that's ticking
over nicely -- I'm pretty happy with it and how it's performing. It's
a 32 bit build, and the machine I'm running it on has a lot of extra
memory.

I assume I'll have to do a 64 bit build to use more than a few gig of
shared buffers. If I do that, though, am I going to have to do a
database dump and reload, or will the disk files be compatible and it
just a matter of shutting down the 32 bit server and firing up the 64
bit one?
--
                Dan

--------------------------------------it's like this-------------------
Dan Sugalski                          even samurai
dan@sidhe.org                         have teddy bears and even
                                       teddy bears get drunk

Dan Sugalski <dan@sidhe.org> writes:
> I assume I'll have to do a 64 bit build to use more than a few gig of
> shared buffers. If I do that, though, am I going to have to do a
> database dump and reload,

Yes, most likely, because you'll have changed MAXALIGN and therefore the
data alignment rules.

You should first ask yourself whether you will get any performance
benefit from having "more than a few gig of shared buffers".  If anyone
has proven such a benefit I haven't seen it.

            regards, tom lane

At 1:01 AM -0500 3/10/07, Tom Lane wrote:
>Dan Sugalski <dan@sidhe.org> writes:
>>  I assume I'll have to do a 64 bit build to use more than a few gig of
>>  shared buffers. If I do that, though, am I going to have to do a
>>  database dump and reload,
>
>Yes, most likely, because you'll have changed MAXALIGN and therefore the
>data alignment rules.
>
>You should first ask yourself whether you will get any performance
>benefit from having "more than a few gig of shared buffers".  If anyone
>has proven such a benefit I haven't seen it.

Possibly it won't. The machine the DB is on sees heavy access to
large files, to the point where parts of the database may get flushed
out of the OS buffer cache. I was working on the (possibly deeply
flawed assumption) that I'd be better off if more of the database was
guaranteed pinned in memory in Postgres' buffer cache -- it wouldn't
necessarily make the peak performance better, but it would make
average performance better, since I'd not have to sometimes hit disk
to read in things that had been evicted from the disk cache.
--
                Dan

--------------------------------------it's like this-------------------
Dan Sugalski                          even samurai
dan@sidhe.org                         have teddy bears and even
                                       teddy bears get drunk

On Sat, Mar 10, 2007 at 08:30:20AM -0500, Dan Sugalski wrote:
> Possibly it won't. The machine the DB is on sees heavy access to
> large files, to the point where parts of the database may get flushed
> out of the OS buffer cache. I was working on the (possibly deeply
> flawed assumption) that I'd be better off if more of the database was
> guaranteed pinned in memory in Postgres' buffer cache -- it wouldn't

Err, is shared memory actually guarenteed to be pinned in memory? I
mean, in Linux it's not as that is a form of DOS attack, pin all memory
by allocating it as SHM.

Now, Solaris may do this differently, but if shared memory can be
swapped out then having lots of it definitly isn't a beniefit.

Have a nice day,
--
Martijn van Oosterhout   <kleptog@svana.org>   http://svana.org/kleptog/
> From each according to his ability. To each according to his ability to litigate.

At 7:47 PM +0100 3/10/07, Martijn van Oosterhout wrote:
>On Sat, Mar 10, 2007 at 08:30:20AM -0500, Dan Sugalski wrote:
>>  Possibly it won't. The machine the DB is on sees heavy access to
>>  large files, to the point where parts of the database may get flushed
>>  out of the OS buffer cache. I was working on the (possibly deeply
>>  flawed assumption) that I'd be better off if more of the database was
>>  guaranteed pinned in memory in Postgres' buffer cache -- it wouldn't
>
>Err, is shared memory actually guarenteed to be pinned in memory? I
>mean, in Linux it's not as that is a form of DOS attack, pin all memory
>by allocating it as SHM.

I'm not worried about swapping -- the box has more than enough memory
that it's not going to swap.

What I am looking at is response times after a relatively long (20-30
minute) period of inactivity. When the database is being accessed and
the data files are either in shared buffers or OS file cache,
performance is really snappy. When the database has been ignored for
a while and other things go on the box, postgres' data files get
expunged from the OS caches, and access times for the first few
queries go from milliseconds to seconds. No surprise, certainly, HD
access times being what they are.

I've  got enough memory on the box that I can reasonably map a four
or five gig into postgres and just leave it there. It'll probably not
slow down anything else significantly, but if it means the difference
between having to re-read a page from disk and just hit shared
buffers, then it's worth it in response time.
--
                Dan

--------------------------------------it's like this-------------------
Dan Sugalski                          even samurai
dan@sidhe.org                         have teddy bears and even
                                       teddy bears get drunk