Re: ARP/ND (was RE: Two "shared VLAN" alternative proposals)

 

> -----Original Message-----
> From: Silvano Gai [mailto:sgainuovasystems.com] 
> Sent: Tuesday, April 03, 2007 4:22 PM
> To: Caitlin Bestler; J. R. Rivers; Eric Gray (LO/EUS);
Radia 
> Perlman; rbridgepostel.org
> Subject: ARP/ND (was RE: [rbridge] Two "shared
VLAN" 
> alternative proposals)
> 
> 
> Do we have any quantitatively data about the need of
ARP/ND proxy?
> 
> With an ARP cache of 30 minutes, typical in hosts
today, even 
> with a 100 K hosts in a VLAN we get at most 55 ARP
seconds. 
> Since not all the hosts talk with each other, it is
more 
> typically like 5 ARP/sec.
> 
> The ARP/ND cache on the RBridge must be significantly
shorter 
> than 30 minutes to not increase the amount of obsolete
information.
> 
> Are we reducing from 5 ARP/sec to 3 ARP/sec for 100 K
hosts per VLAN?
> 
> Is this the big optimization for which we care to
create 
> corner cases and potential incompatibilities?
> 

The ARP/ND transaction rate is not the critical issue.

The more critical issue is the end-to-end latency of a
transaction
that requires doing an ARP/ND first. As the subnet gets
larger,
and the number of "L2 Hops" increases, that answer
gets worse
unless you implement a distributed ARP/ND proxy service.





_______________________________________________
rbridge mailing list
rbridgepostel.org
ht
tp://mailman.postel.org/mailman/listinfo/rbridge

I am surprised,
in ARP the dominant time is the software processing on the
two nodes
that goes through a classical interrupt moderation technique
and exceed
multiple milliseconds. The Rbridge latency is probably 1 to
10
microseconds, so even if you have 100 hops, it is still
negligible.

-- Silvano

> -----Original Message-----
> From: Caitlin Bestler [mailto:caitlinbbroadcom.com]
> Sent: Tuesday, April 03, 2007 4:37 PM
> To: Silvano Gai; J. R. Rivers; Eric Gray (LO/EUS);
Radia Perlman;
> rbridgepostel.org
> Subject: RE: ARP/ND (was RE: [rbridge] Two "shared
VLAN" alternative
> proposals)
> 
> 
> 
> > -----Original Message-----
> > From: Silvano Gai [mailto:sgainuovasystems.com]
> > Sent: Tuesday, April 03, 2007 4:22 PM
> > To: Caitlin Bestler; J. R. Rivers; Eric Gray
(LO/EUS); Radia
> > Perlman; rbridgepostel.org
> > Subject: ARP/ND (was RE: [rbridge] Two
"shared VLAN"
> > alternative proposals)
> >
> >
> > Do we have any quantitatively data about the need
of ARP/ND proxy?
> >
> > With an ARP cache of 30 minutes, typical in hosts
today, even
> > with a 100 K hosts in a VLAN we get at most 55 ARP
seconds.
> > Since not all the hosts talk with each other, it
is more
> > typically like 5 ARP/sec.
> >
> > The ARP/ND cache on the RBridge must be
significantly shorter
> > than 30 minutes to not increase the amount of
obsolete information.
> >
> > Are we reducing from 5 ARP/sec to 3 ARP/sec for
100 K hosts per
VLAN?
> >
> > Is this the big optimization for which we care to
create
> > corner cases and potential incompatibilities?
> >
> 
> The ARP/ND transaction rate is not the critical issue.
> 
> The more critical issue is the end-to-end latency of a
transaction
> that requires doing an ARP/ND first. As the subnet gets
larger,
> and the number of "L2 Hops" increases, that
answer gets worse
> unless you implement a distributed ARP/ND proxy
service.
> 
> 
> 


_______________________________________________
rbridge mailing list
rbridgepostel.org
ht
tp://mailman.postel.org/mailman/listinfo/rbridge