Thanks for all the responses!  I'll comment on a few:

On Mon, Oct 01, 2018 at 06:26:05PM +0200, Lorenzo Bianconi wrote:
> On Sep 26, Ben Pfaff wrote:
> > A 30-second timeout seems long to me.  After that long, the sender will
> > have retransmitted the packet, probably multiple times.  I wasn't able
> > to figure out how long Linux buffers packets.
> > 
> 
> afaiu Linux neighbor subsystem removes an unresolved entry after a
> given number of attempts (~5). The default value for NEIGH_VAR_RETRANS_TIME
> is 1s, so I guess we can use 5s here as BUFFER_MAP_TIMEOUT?

That sounds reasonable to me.

> > Do you think it makes sense to use the same timestamp for all of the
> > packets buffered for a single IP?  I believe that, currently, no packets
> > will ever expire from a buffer unless either the buffer limit is
> > exceeded or no new packets have been queued for the timeout interval.  I
> > think that would mean that there could be a packet buffered for a total
> > of 30 + 64 seconds.
> 
> Stale timeout was intended for buffered_packets_map entries and not for 
> packets
> buffered for a single IP. However reviewing the code I agree that wait
> so long for a given packet is not useful. On the other hand if we set queue
> depth to 3 or so I guess it will not make a huge difference. Do you agree?
> Anyway we can add per-packet timeout instead of per entry one

I agree that with a small queue depth, it doesn't matter as much, so I
think either place for a timeout is fine.

Thanks,

Ben.
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to