From: E. Richard Glazier (erglazier@hotmail.com)
Date: Wed Sep 04 2002 - 20:55:40 EDT
Sorry for the barrage of questions, but I wanted some clarification on some
of these memory / monitoring issues and I knew this would be the place to
ask. Some of these I'm sure I have the correct understanding of, but there
is some conflicting information out there (including old Tru64 manager
lists), so anything you can provide would be great. I will summarize. TIA.
Tru64 5.1 PK4 on a GS80- 8 cpu, 16GB mem.
--------------------------------------------------------------------
If UBC is shared for filesystem cashing and VM, what is the rest of
available physical memory used for?
When I see free pages (from vmstat, collect, monitor, etc), is that free
pages from UBC and otherwise?
Does the fact UBC shares it's space with VM mean that UBC is for filesystem
cache only, until VM free pages outside UBC are gone (ubc_max_percent is
55%) and it then borrows pages from UBC?
Does "free memory" not include UBC, or is it just free pages anywhere?
Is the shared memory segment(s) hard wired outside of UBC, or is it part of
UBC? How is this reflected in vmstat / vmubc output? Is it part of "active
pages"?
Does "paging" mean page-out operations- i.e if I see 0 "page outs" in
vmstat, does that mean I'm not paging?
What is a page-in operation? Is that just brining a dirty page from the
swap space to the free list?
I have plenty of UBC (the hit rate is around 88%), plenty of free memory
(around 2GB), and 0 page-outs, but I have roughly the same number of
page-ins as I did before I added more memory. Do my vm_page* settings need
to be changed? Are pages getting cleaned because of a time-out and not
because there is a lack of available pages? Are page-ins CPU intensive
operations?
Does a UBC LRU or an "old" VM page that gets put to disk not constitute a
page-out?
Does my swap space need to be at least equal the amount of physical memory
because in theory that many dirty pages could be put to disk (it currently
is less than half, but I'm not paging or swapping)?
Should swap_egar be 0 for best performance, if you have pleanty of free
memory?
Should my UBC hit rate be higher than 88%? Would increasing the
ubc_min_percent (currently 10%) help this? It doesn't look like UBC gets
close to the ubc_min_percent value. Is a hit rate of close to 100%
unrealistic because there are always going to be new i/o fetches for data
that's never been cashed before?
What are granularity hints? Do we need Direct I/O to take advantage of
this? How would I set this up having an unpartitioned GS? Am I better of
to use ssm_threshhold?
_________________________________________________________________
Join the world’s largest e-mail service with MSN Hotmail.
http://www.hotmail.com
This archive was generated by hypermail 2.1.7 : Sat Apr 12 2008 - 10:48:51 EDT