topic Re: pseudo-swap question in Operating System - HP-UX

pseudo-swap question

Chuck Ciesinski — Mon, 11 Jul 2005 14:00:34 GMT

To all,

Please forgive me if the answer is already available as I just got tasked to find out if a patch has been released to fix known problem
# KBRC00015724.

Thanks in advance for any and all assisistance...

Chuck Ciesinski

Re: pseudo-swap question

Patrick Wallek — Mon, 11 Jul 2005 14:06:34 GMT

Umm...What's the problem? The Doc ID you give just answers the question "What is pseudo swap?"

What problem are you trying to solve?

If you are wanting to use pseudo-swap, just turn on (set to 1) the kernel parameter swapmem_on, recompile the kernel and reboot.

If you are looking for something else, please provide more information.

Re: pseudo-swap question

Chuck Ciesinski — Mon, 11 Jul 2005 14:46:27 GMT

Patrick,

We have a system which one day jumped from 45% utilization of configured memory to 92% utilization and the percentage has stayed at the high utilization level. We are exploring the excessive potential utilization of pseudo-swap and potential recovery processing. The system is an rp8400 with 11.11 v1 ( December 2004 patch level) with 24 GB of memory.

Chuck

Re: pseudo-swap question

D Block 2 — Mon, 11 Jul 2005 14:53:30 GMT

chuck,

what is your kernel parameters:

# kmtune |grep -i dbc
dbc_max_pct 3 - 3
dbc_min_pct 1 - 1

Re: pseudo-swap question

A. Clay Stephenson — Mon, 11 Jul 2005 15:01:57 GMT

Pseudoswap isn't swap at all; it's simply bookkeeping and doesn't use memory. Do a swapinfo -t, the memory row lists the value for pseudoswap but it's really simply space used by processes. A very common cause of the problem you observe is unused/unattached shared memory segments. Do an ipcs -ma and note any large shmid's that have NATTCH = 0. These might possibly be safe to remove; nattach = 0 is a necessary but not necessarily sufficient condition for safe removal of a shmid via ipcrm. Shared memory segments can easily be left dangling if kill -9 has been used. You could also have very large buffer cache settings. On boxers with your amount of memory it makes much more sense to set bufpages to a non-zero value because even small percentages of dbc_xxx_pct can be quite large with 24GB of memory. In general, 800-1600MB for buffer cache is very, very generous so set bufpages to a equal to the number of desired 4KB chunks.

Re: pseudo-swap question

Ted Buis — Tue, 12 Jul 2005 08:30:12 GMT

Chuck,

Have you rebooted since the jump in memory utilization? Any zombies?

I can't find KBRC00015724.