From: Holger.VanKoll@SWISSCOM.COM
Date: Wed May 12 2004 - 10:05:10 EDT
Hello,
I just openen this pmr with ibm; maybe someone has some feedback?
Regards, Holger
This system has 102 filesystems.
While setting this system up I had to increase numfsbufs (vmtune -b) to
10000 until I got no more (or few) fsbufwaitcounts.
Everything worked fine until I had to create more filesystems later.
After creating 12 more filesystems, I could not mount the 12th
filesystem anymore:
mount: 0506-324 Cannot mount /dev/p00i2n0012 on /data/db2p00i2/NODE0012:
There is not enough memory available now.
I assumed some shortage of pbufs/fsbufs and looked at the
kernel-segment:
sbe12169 # svmon -S 0
Vsid Esid Type Description LPage Inuse Pin Pgsp
Virtual
0 - work kernel seg - 45270 25240 804
45270
Then, for testing, I decreased numfsbufs to 100:
sbe12169 # /usr/samples/kernel/vmtune -b 100
and unmounted the 11 new filesystems and looked at kernel-mem again:
sbe12169 # svmon -S 0
Vsid Esid Type Description LPage Inuse Pin Pgsp
Virtual
0 - work kernel seg - 45270 17944 804
45270
Pinned memory got down from 25240 to 17944 pages.
After remounting the 11 filesystems it increased only to 17979 pages.
Then I also could mount the 12th filesystem.
So, some limitation of kernel pinned memory prevents me from mounting
more filesystems with numfsbufs=10000.
I need to get as close as possible to 10000.
The problem is that changes to numfsbufs only get effective when the
filesystems are unmounted. As the other (102) filesystems are
productive, I cannot mount/umount them for try-and-error.
So, please tell me the limits of kernel/vmm-buffering or where to read
about it.
Regards,
Holger
This archive was generated by hypermail 2.1.7 : Wed Apr 09 2008 - 22:17:54 EDT