Hi All,
Replying to my own posting.... A colleague found this posting which
sound like it should fix our problem. Thanks for the feedback I received.
http://www.uwsg.indiana.edu/hypermail/linux/kernel/9911.0/0094.html
regards,
Stephen.
On Mon, 26 Jun 2000, Stephen J. Gowdy wrote:
> Hi All,
> I was wondering if anyone could shed any light on this problem
> we're having. Our experiment (BaBar - http://www.slac.stanford.edu/BFROOT/
> ) started to use Linux this year
> (http://chep2000.pd.infn.it/paper/pap-e309.pdf ) as one of our supported
> platforms. We take a _lot_ of data, we're getting close to 1TB/day. At the
> moment most of this is handled using Solaris 2.6 (with various patches to
> support large files) at SLAC. We use an object database and allow the
> files it uses to grow to ~10GB each. One of the driving reasons for such
> large files is that our database is limited to have 64k files accessible
> from it.
> However, to understand this data we do a lot of simulation. As
> this is done all over the world we limit the size of these files to
> 2GB. This was originally due to the Solaris 2.6 limitations but also now
> Linux.
> That isn't the problem we're worried about today. The first site
> to attempt to setup a production cluster using Linux machines fails to
> utilise the full CPU available due to the server machine beginning to
> swap. The observation is that the jobs start off running well. If you
> watch top on the server machine (which has an IDE RAID array) which runs a
> database server process (a page server) the buffer grows to use most of
> the physical memory in the machine. The machine has 512MB and the buff
> grows to 480MB. Once this happens the machine begins to swap and the
> client's CPU utilisation drops dramatically.
> We've found the /proc/sys/vm/buffermem file which sounds like it
> should do what we want (ie limit the amount of memory the buffer will
> use). However, as I'm sure many of you know, it seems these numbers are
> not actually used. I hunted through the kernel source (2.4.0-test2) to see
> if I could find anywhere that used them and failed. There is one macro
> which uses the min_percent portion but that macro doesn't seem to be used.
> The kernel on the server machine is 2.2.13 at the moment. Does
> anyone know a way to limit this buffer size? Or is it a function of file
> sizes that are opened (which will grow to n*2GB)? Could this be an
> application problem or only kernel? Any other advice?
> Thanks for any help or advice you can offer.
>
> regards,
>
> Stephen.
>
>
-- /------------------------------+-=-=-=-=-+-------------------------\ |Stephen J. Gowdy |A4000/040| Mail Stop 50A-2160, LBL, | |http://www.ph.ed.ac.uk/~gowdy/ | 1GB HD| 1 Cyclotron Rd, Berkeley,| | |20MB RAM| CA 94720, USA | |InterNet: SGowdy@lbl.gov |3.4xCDROM| Tel: +1 510 495 2796 | \------------------------------+-=-=-=-=-+-------------------------/- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon Jun 26 2000 - 21:00:10 EST