[OpenAFS] Network becomes terribly slow when cache manager flushes updates over xDSL

Goulou fredericg_99@yahoo.fr
Tue, 7 Jul 2009 19:08:36 +0200 (CEST)


> Sure it is.

I actually knew that (fortunately...), thus couldn't imagine I was alone in=
 this situation...

> If there is significant packet loss on the connection, it will go
> slower still.

I agree with you, I don't think my connection has that many losses though.

>  (a) the cache is larger than the working set size of the files
>      being used so that thrashing does not occur

OK on this

>  (b) the operations being performed are mostly reads

That is the case in general, I don't think I was clear on when the problem =
happened : actually, it is not necessarily when one is using Firefox that i=
t hangs (especially the 3.5, since they reduced the syncs on disk, at last.=
..), but when, for example, the user copies a file (whatever size >~ 1Mb) f=
rom any USB stick to any afs directory. Everything else (including FF) suff=
ers from a round trip >=3D1000ms.
That somewhat surprised me, because I thought that by setting storebehind t=
o 20000kb, I should not notice the slow down below 20Mb : when the user cop=
ies a file to its home directory, it looks quite like local copy, and, afte=
r that, the afs cache manager ensures that in the following minute(s) every=
thing is synced again to the server...

> Programs such as firefox that are performing large numbers of network
> operations of their own are going to be competing with the file
> system
> for the same resources.  Firefox in fact is going to be performing a
> write to its web cache for every read from a web server.   Storing
> such temporary data in network storage over a constrained pipe is not
> going to work well.  You should configure firefox to always use
> temporary local disk space for its cache.

Yep, it's already done (turning off local cache). I think I'll try to put s=
ome (unimportant) directories like firefox cache, ".cache", ".thumbnails" d=
irectory and everythink cache-related to local disk (usually by making a li=
nk to /local-home/userXX/YYY, and by creating those local directories to lo=
cal disks). But it wouldn't solve the lag problem when copying...

Thanks for this advice, I'll keep you posted of my evolutions...

Fr=C3=A9d=C3=A9ric Grelot.

>
> Jeffrey Altman
>
>
> Goulou wrote:
> > Nobody has any idea about this? Isn't AFS designed to be used over
> "normal" internet links? Or am I the only one with this slowing-down
> problem?
> >
> > I'll be really happy if anybody would help me...
> >
> >
> > Fr=C3=A9d=C3=A9ric Grelot.
> >
> >
> > ----- "Goulou" <fredericg_99@yahoo.fr> a =C3=A9crit :
> >
> >> Hi all,
> >>
> >> I'm trying to give my users an access to their home directory over
> the
> >> Internet, so they can connect to there account from their homes
> >> (place, not directory...).
> >> Unfortunately, my tests show that when a large update occurs to
> the
> >> disk (for example, the user download a file and puts it in any
> >> directory, or makes a copy from USB to home dir), the connection
> >> always becomes terribly slow (ping>1000ms), and any other request
> >> takes ages to return (thus, if the user uses firefox for example,
> >> since it makes quite a lot of disk access to the ".mozilla"
> directory,
> >> its interface freezes until the update if over and ping becomes
> normal
> >> again, i.e 60ms)
> >>
> >> I managed to optimize this by setting a chunksize of 8kb (smaller
> >> seems to fail, because my cache is between about 500Mo-2Go, and
> that
> >> makes too many files for the cachemanager I think), and by adding
> an
> >> "fs st -allfiles 20000" in afs.conf like this:
> >> afs_server_prefs() { fs storebehind -allfiles 20000 }
> >> It is slighly better with these settings : I still get ping
> >1000ms,
> >> but less often...
> >>
> >> Is there any ways to enhance this again, or should I set some QoS
> on
> >> the system to prevent afs daemon to send to many packets at the
> same
> >> time (a solution that I'd rather avoid, since I'll not have full
> >> control over the client's systems...)?
> >>
> >> Tanks a lot for your advices!
> >>
> >> Fr=C3=A9d=C3=A9ric.
> >> _______________________________________________
> >> OpenAFS-info mailing list
> >> OpenAFS-info@openafs.org
> >> https://lists.openafs.org/mailman/listinfo/openafs-info
> > _______________________________________________
> > OpenAFS-info mailing list
> > OpenAFS-info@openafs.org
> > https://lists.openafs.org/mailman/listinfo/openafs-info
> >