Re: max_files_per_process limit

On Mon, Nov 10, 2008 at 5:24 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> "=?ISO-8859-1?Q?Dilek_K=FC=E7=FCk?=" <dilekkucuk(at)gmail(dot)com> writes:
> > We have a database of about 62000 tables (about 2000 tablespaces) with an
> > index on each table. Postgresql version is 8.1.
>
> You should probably rethink that schema. A lot of similar tables can be
> folded into one table with an extra key column. Also, where did you get
> the idea that 2000 tablespaces would be a good thing? There's really no
> point in more than one per spindle or filesystem.
>
> > Although after the initial inserts to about 32000 tables the subsequent
> > inserts are considerable fast, subsequent inserts to more than 32000
> tables
> > are very slow.
>
> This has probably got more to do with inefficiencies of your filesystem
> than anything else --- did you pick one that scales well to lots of
> files per directory?

The database is working on FreeBSD 6.3 with UFS file system. It has 32 GB of
RAM with 2 quadcore Intel Xeon 2.66 GHz processor, and about 11 TB of RAID5
storage.

>
>
> > This seems to be due to the datatype (integer) of max_files_per_process
> > option in the postgres.conf file which is used to set the maximum number
> of
> > open file descriptors.
>
> It's not so much the datatype of max_files_per_process as the datatype
> of kernel file descriptors that's the limitation ...

We do not get any system messages related to the kernel file descriptor
limit (like file: table is full) yet we will work again on both the database
schema (tablespaces etc.) and system kernel variables.

Thanks,
Dilek Küçük

>
>
> regards, tom lane
>