On Wed, 2002-06-19 at 06:22, KELEMEN Peter wrote:
>
> [ For those who haven't followed #xfs: after plotting the stats
> and finding correlation with xs_ig_attrchg counters, Steve came
> up with a theory of excessive utime() activity causing the high
> number of empty transactions. ]
>
> Your theory was correct; after six days of monitoring, the culprit
> application was found to be ncftp. You can reproduce the effect
> by simply downloading a big file, and it will boost your empty
> transaction counter. Snippet from the strace output:
>
>
> There was no definite opinion if we need these transactions at
> all, and you mentioned that eliminating them would be a more
> coplex change. Awaiting your comment.
>
> Peter
Still pondering this one, I suspect the correct answer is we need both,
XFS has a wsync mount option which makes transactions fairly
synchronous. In that case we should be logging the inode, in the
normal case, I think the transaction should be removed. The
transaction would not be written out to disk until something
else pushes it out, which would often be periodic sync activity, which
would be just as happy to push the inode itself out.
Steve
--
Steve Lord voice: +1-651-683-3511
Principal Engineer, Filesystem Software email: lord@xxxxxxx
|