Please don't reply to lustre-devel. Instead, comment in Bugzilla by using the
following link:
https://bugzilla.lustre.org/show_bug.cgi?id=11089
(In reply to comment #9)
> > I just got permission to view this thread. This patch appears to defeat the
> > purpose of create a NID in the proc tree. The reason we need NID is to save
> > all
> > stats even after an export is gone, i.e., after a linux client does an
> > umount or
> > a liblustre client exits.
> >
>
> If you save the stats after the client disappears, you are effectively losing
> about 8k every time a client restarts. This rapidly will cause problems on
> large systems. Or are you talking about re-using the same stats based on the
> nid? If so, then this fails if you have two clients mounted on the same node
> (a
> fully-supported situation), where they will overwrite each other.
> I think I need to understand your goals before trying to work out any
> implementation.
The proc is organized as this
+- nid --> stats...
|
+ exp#1 --> stats...
|
+ exp#2 --> stats...
So it would avoid both situation you mentioned earlier with a little more memory
consumed for each NID. We considered it is worthwhile unless concrete numbers on
its weakness of scaling.
_______________________________________________
Lustre-devel mailing list
[email protected]
https://mail.clusterfs.com/mailman/listinfo/lustre-devel
