Greetings -
I just identified, via my various monitoring utilities and dbx, the following:
AlphaServer 8400 5/400 (4CPU, 6GB RAM, 6GB deferred-mode swap,
DU 3.2G (rev. 62), NSR v4.2B)
While the following was occurring, we had 2400 processes on the system and
450 users accessing Oracle databases, with just a very slight response time
pause.
Started up NSR v4.2B after installing it over NSR v3.2 according to the
installation instructions.
While nsrmmdbd cross checked the 2 indices (sizes 2M and 175M) it grabbed
upwards of 1600M of resident set memory, causing the following:
Free memory dropped down to and stayed at *9* pages (we have 772587 VM
managed pages, or 6GB RAM).
UBC min percent is set to 10% (77258 pages).
UBC borrow percent is set to 20% (154516 pages).
UBC max percent is set to 80% (618071 pages).
Observed UBC pages dropped down to 55992 pages (170MB less than UBC min%)!
Our 6GB of swap space dropped steadily to 60% free until we shutdown NSR.
At the moment NSR shutdown completed our free memory suddenly grew by 300,000
pages (2.4G).
I've looked through much of the VM subsystem code, and found two references
in vfs_ubc's ubc_page_stealer() (v3.2 source tree) that return from the
ubc_page_stealer() function if the number of UBC pages is less than or
equal to the number of UBC min pages (line 1065), and if that doesn't return,
then at line 1093, there is another abort if either UBC pages is less than
UBC min pages, or UBC pages is less than UBC borrow limit and the number
of pages requested to steal is not zero.
But I just saw that the UBC did drop below ubc_minpages by 21266 pages!
Before I go and try to understand the various references to these sections of
code, does anybody have information about this?
We've verified NSR is the cause - we started NSR again to verify it was the
cause, and the same results occurred.
Have others seen this with NSR v4.2B?
Summary pending responses.
Randy Hayman
haymanr_at_icefog.alaska.edu
Received on Thu May 22 1997 - 22:22:34 NZST