[OpenAFS-devel] Patch to disable fsync()ing
Dale Ghent
daleg@umbc.edu
Fri, 16 Mar 2007 16:48:03 -0400
On Mar 16, 2007, at 1:22 PM, Jim Rees wrote:
> The only reason to fsync would be to enforce a particular ordering
> on the
> data written to disk.
For POSIX write semantics, yes?
> But there is no question afs is way too aggressive at
> syncing.
The main reason behind Rob's fsync patch is because we're using ZFS
on our AFS servers, all 8 of them.
ZFS's current behavior (in the future this will be dynamically
alterable) when dealing with a fsync is that it will flush all
pending IOs in the ZFS intent log (aka ZIL, a write cache) to disk,
and this holds up subsequent write IOs until this is completed.
Because we still run our per-user mail store out of ZFS, fsyncs are
quite expensive because of the sheer amount of data (incoming mail)
being written.
/dale
--
Dale Ghent
UNIX and Storage Systems Specialist
UMBC - Office of Information Technology
ENG 201 - 410-443-1705