Re: NFS alternative
On Mon, 16 Jul 2001, D-Man wrote:
> On Mon, Jul 16, 2001 at 12:45:22PM -0500, Dimitri Maziuk wrote:
> | * D-Man (dsh8290@rit.edu) spake thusly:
> | ...
> | > Ok, that makes sense. How about if probability leaves us behind
> | > and a packet is lost? Does NFS provide any way to correct for
> | > that or will your filesystem be hosed?
> |
> | Thankfully, I forget the details[0]. From experience, no, it won't
> | be exactly hosed: you'll end up with a .nfs004950384672385721380937
> | file that will grow and eventually fill up the partition... nothing
> | an rm -rf / won't fix. And then there's negative cookies and stale
> | mounts that require a reboot on most unices I've seen...
>
> Ok. It sounds like it would still result in data loss :-(.
Hmm, I'm not an NFS expert but I'll play one on the mailing-list for
you ;-) Please, if there are experts out there, correct me if I'm wrong.
AFAIU, NFS has its own mechanism to recover from lost packets. So it
won't be a problem if a packet is lost. Similarly I believe NFS RPCs
cannot span UDP packets, so there is no chance that a lost packet would
change the meaning of an RPC. The RPC will be lost, pure and simple, and
NFS will have to reissue it or something similar. So I don't think
packet loss is an issue.
What NFS is 'lacking' is congestion control, as in the TCP slow star
and exponential back-off. This means NFS will blast UDP packets as fast
as it cans with no regard for other trafic. This is not really an issue
on a lan and actually had a performance at a time (I think). But if you
go over multiple links, then you may saturate a slower link, causing the
router that is just before it to start dropping packets. Especially if
multiple streams converge there. And once you start dropping packets
performance degrades very significantly. I believe that's why NFS is bad
if there are multiple hops (I get it from a very reliable source that
this is also why it's very bad if the traffic will go over ATM, you need
buffering/traffic shaping).
--
Francois Gouget fgouget@free.fr http://fgouget.free.fr/
Any sufficiently advanced bug is indistinguishable from a feature.
-- from some indian guy
Reply to: