git: kernel - Change the vfs.vm_cycle_point default from 23 to 4
commit d300946ff7ce1edbb890f3c52938535dc1949b90
Author: Matthew Dillon <dillon@apollo.backplane.com>
Date: Mon Mar 22 22:44:33 2010 -0700
kernel - Change the vfs.vm_cycle_point default from 23 to 40
* Large volumes of filesystem activity can cause a lot of data cycling
to occur and we generally want this cycling to occur within the inactive
queue where it doesn't push-out idle but strategic pages related to
resident programs in the active queue.
A larger vm_cycle_point value will tend to cause VM pages related to
released buffer cache buffers to be cycled back into the inactive queue,
leaving the pages in the active queue alone. A smaller vm_cycle_point
value will move the more often-used buffer cache buffer pages to the
active queue when the buffer is released.
* Unfortunately there are typically a lot of VM pages related to resident
programs such as xterms and databases which can remain idle for long
periods of time. Excessive cycling of buffer cache pages to the
active queue can cause these long-idle pages to be pushed out to
swap and/or freed.
This results in an annoyance when these processes go active again,
particular ssh, xterm, and other interactive programs. These are
not necessarily going to perform badly once the pages are paged back
in, however.
* It has been found that certain large-scale scans of the filesystem such
as the HAMMER reblocker and overnight cron jobs could push active pages
out unnecessarily. Raising vfs.vm_cycle_point to 40 essentially fixes
this problem but at the cost of potentially starving the 'file cache'
for data sets which would otherwise fit into memory (and for which the
sysop might desire to fit into memory). So YMMV.
Testing-by: Francois Tigeot <ftigeot@wolfpond.org>
Summary of changes:
sys/kern/vfs_bio.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/d300946ff7ce1edbb890f3c52938535dc1949b90
--
DragonFly BSD source repository