RE: [Simple] Harmonizing MSRP and SIMS (long)

[Adam Roach <adam at dynamicsoft.com>]

Orit Levin [mailto:oritl@microsoft.com] wrote:

> Nope. It means to have the ability to work in
> a mode when the response from the next hop is not
> being sent (not per message neither per a bulk
> of messages). If something does go wrong, NACK is
> sent (physically from the next hop, logically -
> not necessarily). 
> 
> Otherwise, MSRP will not scale for BIG systems.

I'm not certain I buy this assertion. Barriers to
scalability come into play when you have processing
that grows faster than linearly with the size of the
system. Transaction-level acknowledgements are a
simple linear increase over unacknowledged messages.
If necessary, acknowledgements can be made very, very
small (on the order of a dozen bytes or so), which
makes them a very small percentage increase over
unacknowledged operation.

The problems with unacknowledged operation over
TCP stem largely from the fact that a lost
connection can take a rather long time to percolate
up to the application. Once such a failure is
detected, the application needs to try to re-send
any potentially lost messages. With acknowledged
transactions, this is easy: any unacknowledged
transactions are retransmitted once an alternate
connection is established (with the corollary
that buffered messages can be discarded as soon
as they are acknowledged).

With unacknowledged messages, the application
essentially has to keep a theoretically infinite
buffer (which, in practice, means "some very large
number, multiplied by the link speed, and then
multiplied by the next hop latency") to do anything
sensible when a TCP connection glitches. Now *that*
will have trouble scaling.

> IM is more like RTP/RTCP for media than SIP for
> signaling.

Not in most important respects. With RTP, loss of a
packet cannot be recovered. By the time that the sender
could detect loss, it's already too late to do anything
for that specific loss. IM is very different in that
respect.

RR packets in RTCP provide aggregate statistics only
because specific RTP packets aren't interesting.
Individually, they're disposable. If one were to design
a similar RR packet for a protocol in which each "packet"
were valuable (like IM), then these records would
undoubtedly include acknowledgement of each received
message.

/a

_______________________________________________
Simple mailing list
Simple@ietf.org
https://www1.ietf.org/mailman/listinfo/simple