Steve,
The disk under the hood is IBM raid array (sharks?). We're seeing them
through a qlogic driver to the SAN.
If you've got any tuning params. or patches for me to try, let me know.
We like the performance of xfs (what we were able to measure), but need to
pick a filesystem soon. This will be the filesystem behind our 512 host
linux cluster at NCSA.
I downloaded the development tree from your cvs server, but was not able
to "make bzImage" on linux 2.4.4:
gcc -I/usr/include -ldb1 aicasm_gram.c aicasm_scan.c aicasm.c
aicasm_symbol.c -o aicasm aicasm_symbol.c:39: db1/db.h: No such file or
directory
make[5]: *** [aicasm] Error 1
So we're still on linux 2.4.2. I down-graded the system for testing xfs
because your web site made it look like 2.4.2 was the development
platform. If you've got a working 2.4.4 version I'd like to try it out
(that's what the rest of the cluster is running).
-Galen
On Tue, 22 May 2001, Steve Lord wrote:
>
> >
> > Steve,
> >
> > I ran "iozone -s 4000m" again with a new kernel (your attached patch).
> > I'm getting the same errors and have attached a slab.out.gz file showing
> > slabinfo every 1s during the test. The __alloc_pages: error shows up a
> > few seconds into the test. buffer_head values seem to ramp up
> > quickly with iozone...
> >
> > -Galen
> >
>
> One more question, are you running on top of md/lvm? I see you have a lot of
> disk hardware connected to the box, just wondering if iozone is getting
> to access lots of it, or just one device out of all the ones you have.
>
> The fundamental issue appears to be dirtying memory way faster than it can
> be cleaned, and not throttling the dirtying of memory correctly. You have
> 12000 pages worth of buffer heads allocated, which may be a bit excessive!
>
> Steve
>
>
> Steve
>
>
--
+
Galen Arnold, system engineer--systems group arnoldg@xxxxxxxxxxxxx
National Center for Supercomputing Applications (217) 244-3473
152 Computer Applications Bldg., 605 E. Spfld. Ave., Champaign, IL 61820
|