On Sat, 11 Aug 2001 at 23:51, Seth Mos wrote:
> Weird, I expected the data to get written first and then the metadata.
> If that would happen you would have no file, or in our case, the old
Does anybody at all remember my post regarding bdflush settings? I repeat
two contents of /proc/sys/vm/bdflush that seem to be related:
age_buffer = 3000
age_super = 60
These are the defaults, and are in jiffies. Here is a quote from
-----[ proc.txt ]-----
age_buffer and age_super
Finally, the age_buffer and age_super parameters govern the maximum time
Linux waits before writing out a dirty buffer to disk. The value is
expressed in jiffies (clockticks), the number of jiffies per second is
100. Age_buffer is the maximum age for data blocks, while age_super is
for filesystems meta data.
-----[ end proc.txt ]-----
You cannot set age_buffer to be < 100, I don't know why (didn't find
documentation for this but learned experimentally that it won't accept
anything < 100), but still, 100 jiffies is 1 second.
Now my still unanswered question: isn't this all that we need to make sure
data gets on to disk in an idle system? I've already set my server's
/proc/sys/vm/bdflush to have an age_buffer = 100, although I haven't done
similar tests pulling the plug because I just don't do that with the
server. I will find time to simulate such tests with the two different
bdflush settings on a non-production machine.
I _think_ that setting bdflush with age_buffer == 100 will give us "the
best of both worlds". On an active system you still don't get O_SYNC
performance as you still get to use the buffer, and yet on an idle system
you don't wait 3000 jiffies to flush dirty buffers to disk.
Perhaps someone more into testing can do this very simple comparison,
although I hope to send the list follow up information soon.
Federico Sevilla III :: jijo@xxxxxxxxxxxxxxxxxxxx
Network Administrator :: The Leather Collection, Inc.
GnuPG Key: <http://www.leathercollection.ph/jijo.gpg>