> > On (04/30/10 00:32), Sven Axelsson wrote:
> >> The only way to stop the arp requests is to manually delete the arp entry.

See CR 6948860: 
6948860 nce_timer() checks for return value from arp_request() are broken

On (04/30/10 07:40), James Carlson wrote:
> This 2-packets-per-second state sounds like a bug to me.
> 
> Just because you have traffic to send is not a good reason to send yet
> another ARP request -- it leads to ARP broadcast flooding, which will
> cause the switches that Oracle builds into its rack systems to start
> dropping traffic, among other problems.
> 
> Doing the first retry in 500 milliseconds might not be a bad idea, but I
> think there should be a back-off algorithm of some sort here.
> 
> (Traditional ARP, if I remember correctly, would just not try again
> until the entry timed out, so you ended up with something like one
> packet every 20 seconds.  That's probably not enough for modern
> networks, but 2 per second might be too many.)

yes, after the fix for CR 6948860, we reverify the ARP information for
the first packet (after a long period of silence) that's sent out, and
while the conversation is succesfully active, we re-verify the information
every 30 s +/- some fuzz.

So the switches are safe! :-)

--Sowmini



_______________________________________________
networking-discuss mailing list
[email protected]

Reply via email to