RE: MOPO and the 'ping' probes

> -----Original Message-----
> From: Bill Sommerfeld [mailto:[email protected]]
> Sent: Tuesday, March 08, 2005 8:22 AM
> To: Timothy Liu
> Cc: [email protected]; Eric Vyncke; [email protected]
> Subject: RE: [Mobike] MOPO and the 'ping' probes
> 
> On Tue, 2005-03-08 at 10:06, Timothy Liu wrote:
> > There is an approach base on same idea, I am not sure if this has
been
> > brought up before. Why not send DPD with the same sequence number
> > (simultaneously?) to all the paths that the peer wants to test? From
the
> > other side it can treat is as retransmission and resend the DPD
response
> > through the different paths. The retransmission policy on the
requester
> > side is only to retry along the primary path.
> 
> yes, but you also need to handle the case when the sender didn't have
> enough
> window for even one request because all window slots were occupied by
> messages
> which hadn't gone through because the preferred path was broken..

Good point.

> 
> One solution would be to use a real request instead of a DPD ping
> when you actually have something real to send.
> 
> I'd think simultaneous probes would be considered unfriendly to the
> network;
> non-simulataneous (delay of a large fraction of the RTT or so between
each
> probe)
> might be nearly as good at finding a path without generating nearly as
> much redundant
> traffic...
> 
> 

How about this for a path test: cycle through the non-primary paths you
want to test, use the last request message sent to the peer. (Does not
matter whether this request has already got a reply or not). To the peer
it is going to look like a new request, which it is going to process and
respond anyway, or a retransmission, for which it is going to retransmit
the last response.

The exception is if the last message is PATH_CHANGE. Here to avoid out
of order arrival and change to a wrong path, you just have to wait for
it to complete. If PATH_CHANGE fails, we would need to tear down the SA.
But to recover from it we may need window size > 1 anyway.