According to all the Tru64 documentation I could find, when the system needs
to reclaim some memory for the free list, it reclaims it from the UBC if the
percentage of memory in use by the UBC is greater than ubc-borrowpercent,
which is clearly the case on our system (see below). So you'd expect page
outs to be nil because it won't page out any vmm process pages until its run
out of borrowed UBC pages to reclaim.
However, we are still getting paging out activity (up to 15 pages / sec) as
seen in vmstat and collect. Theoretically this shouldn't be the case, and I
was wondering if anybody had any clues as to why it is still paging out.
Basically, we want to increase the size of the database server's cache to
get more hits in there (rather than AdvFS or UBC hits), but are unwilling to
increase memory usage when the system is already paging out. We are also
going to reduce the size of the AdvFS buffer, but want to sort this out
first.
The system in question is an 8200 with 8 Gb memory. It is running as a
Progress database server.
The following relevant sysconfig parameters are set:
ubc-borrowpercent = 20
ubc-minpercent=10
ubc-maxpercent=100
vm-page-free-target = 1024
vm-page-prewrite-target = 2048
vmstat -P gives:
free pages = 1151 (8 Mb)
active pages = 52589 (410 Mb)
inactive pages = 92117 (719 Mb)
wired pages = 103891 (811 Mb)
ubc pages = 779561 (6090 Mb) ****** = 75 % *******
Total = 1029309 (8041 Mb)
The size of the AdvFS buffer cache is ~570 Mb (7%) of physical memory
(included in "wired pages" above).
A second but much less important question is what does the "page in" value
include? Obviously it includes the page-in on demand of program text, but
does it also include paging in pages from standard files in an AdvFS
filesystem?
Thanks in advance,
Ben Stanley.
ben.stanley_at_gbst.com
Received on Mon Sep 18 2000 - 05:52:54 NZST