xfsaild centisecs /proc wakeup control?
I was looking at powertop again recently, and noticed xfsbufd and xfsaild at the top of the wakeup list. xfsbufd can be tuned with /proc/sys/fs/xfs/xfsbufd_centisecs along with 3 other time-related tunables. However, I don't see one for xfsaild. Is that an oversight?
Quoted 4 lines · Linda A. Walsh
I was looking at powertop again recently, and noticed xfsbufd and xfsaild at the top of the wakeup list.
xfsbufd can be tuned with/proc/sys/fs/xfs/xfsbufd_centisecsalong with 3 other time-related tunables. However, I don't see one for xfsaild. Is that an oversight?
On what kernel?
As far as the wakeups go, I thought this was fixed; see for example:
TAKE 977545 - xfsaild causing too many wakeups
on 2.6.27-rc8 with xfs mounted, and powertop 1.9, I don't see xfs up very high at all:
Top causes for wakeups:
29.4% ( 15.0) <interrupt> : pata_sil680
9.4% ( 4.8) <interrupt> : eth0
7.8% ( 4.0) iscsid : __mod_timer (process_timeout)
...
2.0% ( 1.0) xfsbufd : __mod_timer (process_timeout) 2.0% ( 1.0) xfsaild : __mod_timer (process_timeout)
-Eric
Quoted 5 lines · Linda A. Walsh
I was looking at powertop again recently, and noticed xfsbufd and xfsaild at the top of the wakeup list. xfsbufd can be tuned with
/proc/sys/fs/xfs/xfsbufd_centisecsalong with 3 other time-related tunables. However, I don't see one for xfsaild. Is that an oversight?
No. The xfsaild is currently operating with watchdog functionality to catch AIL push stalls due to a know problem that this new patch:
http://oss.sgi.com/archives/xfs/2008-10/msg00119.html
fixes.
After this has been included for a while (say a release), I plan to remove the watchdog feature of the xfsaild so the once-per-second-per-XFS-filesystem wakeup will go away then.
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 1 line · Eric Sandeen
On what kernel?
2.6.26.0 (haven't rebooted to 2.6.26.5 yet)
Quoted 2 lines
As far as the wakeups go, I thought this was fixed; see for example:
TAKE 977545 - xfsaild causing too many wakeups
---
It may be fixed in a newer kernel...(referring to David's later comment)..
Just looked at it with powertop 1.10:
Wakeups-from-idle per second : 7.1 interval: 10.0s
Top causes for wakeups:
42.9% ( 6.0) xfsaild : schedule_timeout (process_timeout) 14.3% ( 2.0) <kernel core> : clocksource_register (clocksource_watchdog) 8.6% ( 1.2) xfsbufd : schedule_timeout (process_timeout) 7.1% ( 1.0) ip : tg3_open (tg3_timer)