OK so to expand on my previous smug-ness
Upsides:
* No more signaling nat issues. Literally zero. If you want to be
super-sneaky run your edge over TLS port 443 and most things wont
touch you.
* No retransmission's or registration avalanches. They simply cannot
happen since you need a tcp session first.
* No packet fragmentation issues. Send massive bloated SDP's and never
worry about pruning headers again. If you are doing sip SIMPLE send
mime bodies in messages if you want. Its all good.
* Faster convergence (if you reset the TCP connections to your devices
it will usually trigger an instantaneous proxy advance)
* Real-HA on carrier grade SBC's works just fine and TCP state is
maintained across pairs (Acme, Perimeta etc)
* Never chase lost signaling
Downsides:
* Conventional HA doesnt work so well. Your reg/subscription etc will
all be in the context of a single TCP session (with or without TLS).
This means for that second when you restart your proxy the session
is lost and MUST be re-establised by the client.
* SIP Outbound support, which would basically be the answer here
basically doesn't exist in a usable fashion for reliable dual-reg.
Device support is partial and broken. Its not good. There are
potential solutions but it involves real commitment to this right
now and the gulf of experience between having and not isnt huge.
* Moderately more load since TCP state must be retained, but on modern
hardware this is so trivial its almost not worth mentioning.
* Need to re-learn KPI's for network. The entire signaling profile
changes. Its just a different animal.
* Most of your sniffer-based diagnostic tools become useless (for tls)
since packets wont be readable. This is dodged with an edge that
will feed encrypted traffic to a collector.
Suggestions:
STRONGLY recommend terminating TCP/TLS at the edge and still running
core in straight UDP with jumbo frames. You dont want a cascde of tcp
session reestablishments
I have a growing SP network today doing this with great success and also
advise my consulting clients to take this path.
On 8/8/2018 12:36 PM, Alex Balashov wrote:
On Wed, Aug 08, 2018 at 12:21:09PM -0700, Carlos Alvarez wrote:
So...who else on the list uses TCP and has any comments about it?
We are not an ITSP and are Polycom-only with a trivial number of
endpoints, but we do use it and it works just fine.
However, we have numerous customers, some of whom use TCP predominantly
for thousands of endpoints. It works just fine.
In terms of downsides:
In addition to a historical lack of (RFC 3261-mandated) support, there
are of course theoretical trade-offs involved in using TCP. There's
more overhead, and connection state to be maintained on the server side,
which of course consumes resources — resources considered trivial
nowadays, but once upon a time, when RFC 3261 was ratified (2002),
perhaps not. As with all things TCP, it can also present a DoS vector if
you don't limit the number of connections somewhere.
The congestion control/end-to-end delay aspects of TCP are certainly not
as important now as they were at a time when the public IP backbone and
was in an entirely different place in its evolution. Also, nowadays the
congestion/windowing algorithms used in TCP can be tweaked to something
more efficient.
I think the most damning thing about using TCP is perceived to be the
relative difficulty of failing over TCP session state to a different
host. UDP does not require connection state, so as long as you have some
means of handling requests in a relatively stateless fashion, things can
just carry on as they did before in the event of an IP takeover without
anyone having to "reconnect". This is one area where the big enterprise
boxes certainly trump the open-source ecosystem, where transparent TCP
failover *for SIP* doesn't really exist, although in my opinion the
whole issue is getting a bit moot with the way cloud infrastructure and
virtualisation networking is evolving.
-- Alex
_______________________________________________
VoiceOps mailing list
VoiceOps@voiceops.org
https://puck.nether.net/mailman/listinfo/voiceops
