[Gluster-users] performance/writebehind behavior

[Richard van der Hoff]

Thanks for your help with this, Anand, and sorry for sitting on it for a 
while...

On 28/07/14 18:57, Anand Avati wrote:
>         Whether flush-behind is enabled or not, close() will guarantee all
>         previous write()s on that fd have been acknowledged by server.
>
>     Thanks Anand. So can you explain why the 'wc' in my example doesn't
>     see all of the data written by the dd?
>
> I'm wondering if it is because of attribute cache. Maybe attribute cache
> (either in fuse or gluster, don't know yet) is not getting invalidated
> for some reason. Try each of the following and check if any of them make
> the test work right:
>
> #1 mount glusterfs with --attribute-cache=0

I couldn't get this to work - can you be clearer about what I need to run?

$ mount -t glusterfs -o attr_timeout=0 ca1.gl:/shared /mnt/shared2
unknown option attr_timeout (ignored)

> #2 disable stat prefetch with : gluster volume set $name
> performance.stat-prefetch off

This did indeed seem to make the test work right. Does that imply there 
is a cache not being invalidated correctly?

Another thing which I've just realised appears to be relevant: we have a 
process which adds an inotify watch for a subdirectory of the 
fuse-mounted path.

Thanks

Richard