SIPit19 Summary

SIPit 19 took place Oct 16-20, 2006 at the University of New Hampshire InterOperability Laboratory (www.iol.unh.edu).

There were 140 attendees from 73 companies visiting from 16 countries present.
There were 79 teams and 90 distinct implementations.

The Interoperability Laboratory did a spectacular job of providing a rock-solid network for us to test on.
(For those who haven't been to a SIPit, it is a particularly intense network torturing environment).

The majority of the spec-arguments during testing centered around how to handle early media and early dialogs.
There was also a non-trivial subset of the implementers that were confused about whether REGISTER and PUBLISH
create dialogs (much of this confusion seems to come from the presence of to-tags in the 200 OK responses to
REGISTER in the examples in 3261). There were a number of interesting questions about edge cases that I will
send to the appropriate IETF lists separately.

We tried something different for collecting data for this report at this SIPit. We utilized a web-based
automated survey tool. As a result, we collected information on more questions than we usually do, so this
report is a bit long. A side-effect is that the accuracy of the information is probably a little lower. Almost
all of what's below is self-reported, and its unavoidable that for any given question an implementer or two
didn't understand, or didn't know the answer. So, with an appropriate level of respect for errors in sampling, here's
what we found:

The roles represented (some implementations act in more than one role):
  36 endpoint
  19 proxy/registrar
   8 standalone proxy
   4 redirect server
   4 gateway 	
  15 automaton UA (voicemail, conference, etc.)
  17 b2bua/sbc
   6 ua w/ signalling but no media
   8 test/monitoring tool

Implementations using each transport for SIP messages:
  UDP 100%
  TCP  82%
  TLS  45% (server auth only)
  TLS  36% (server or mutual auth)
  SCTP  6%
  DTLS  0%

10% of the implementations supported SIP over multicast.
30% supported SIP over IPv6.

70% of the implementations correctly reassembled fragmented UDP.

Proper use of DNS for SIP continues to rise:
  Full RFC3263 use of DNS  59%
  SRV only 		   14%
  A records only	   15%
  no DNS support           12%

Support for various items:
  32% ENUM
  65% rport
  30% multiplexing SIP/STUN
  14% SIGCOMP
  25% RFC4320 fixes
  14% Identity
  30% connect-reuse


14 of the implementations claimed support for outbound. Interoperability around this draft was fairly low, but the implementers are aggressively improving it.

15 implementations claimed support for some version of GRUU. Nothing worked together before code changes at the event. By the end a few teams were getting scenarios to work.

Only 3 implementations were attempting to support the consent framework.

The endpoints implemented these methods:
  100% INVITE and ACK
  100% CANCEL
  100% BYE
  96% REGISTER
  81% OPTIONS
  76% SUBSCRIBE
  80% NOTIFY
  56% PRACK
  52% MESSAGE
  74% INFO
  63% UPDATE
  80% REFER
  41% PUBLISH

The endpoints implemented these extensions:
  67% RFC3891: replaces
  63% RFC4028: session-timer
  17% RFC3327: path
  11% RFC3840: pref
   4% RFC3841: caller-prefs
  26% RFC3323: privacy
   6% RFC4538: target-dialog
   7% RFC4488: norefersub
  56% RFC3262: 100rel
   3% RFC3994: indication of message composition

44% of the endpoints implemented sipping-cc-transfer

When asked about STUN support, the client implementations replied:
   6% I implement all the client requirements of draft-ietf-behave-rfc3489bis
   6% I implement some, but not all, of the client requirements of draft-ietf-behave-rfc3498bis
  13% I implement all of the client requirements of RFC3489
   7% I implement some, but not all, of the client requirements of RFC3489
  59% I do not implement STUN as a client
   9% Other

There are still a large number of endpoints (25%) that do not use symmetric RTP.

There were only a couple of TURN client implementations. We had several STUN servers and 2 TURN servers.  There were only 3 ICE implementations, and only one of those was at the current version. Interoperability was reasonably high, but not seamless. The issues with interoperability were all implementation problems.

I asked the endpoint implementations to characterize their handling of S/MIME:
  15% I break if someone sends me S/MIME
  22% I pretend S/MIME doesn't exist if it shows up
  35% I don't pay attention to S/MIME, but will proxy it or hand it to my application
   7% I pay attention to S/MIME I receive, but don't send any
   2% I don't pay attention to S/MIME I receive, but I do send some
   4% I try to do something useful with S/MIME I receive and send
  15% Other

I asked the same question about multipart mime support:
   7% I break if someone sends me multipart/mime
  20% I pretend multipart/mime doesn't exist if someone sends it to me
  19% I ignore multipart/mime but will proxy it or hand it to my application if it shows up
  15% I try to do something useful with multipart/mime I receive, but I never send it
   4% I ignore multipart/mime that I receive, but I try to do something useful with multipart/mime I send
  22% I try to do something useful with multipart/mime I send and receive
  13% Other

48% of the endpoint implementations claimed to correctly handle merged requests.

Here is how the endpoints said they handled receiving 200 OKs from more than one branch of a forked INVITE:
  54% I send BYEs to all but one branch
   6% I use all of them (perhaps mixing the different media streams locally)
  16% I don't handle this case gracefully
  11% Other

Here is a sample of how endpoint implementors replied when asked how they handled early media from more than one leg:
  *	We allow multiple RTP streams with an affinity to the last one.
  *	First Media received is played until 200.
  *	The first 183 will be honored in case of the UAC. The rest will be dropped.
  *	Allow media from only negotiated address. Ignore media until negotiated (offer-answer exchanged).
  *	Listen to most recently started stream.
  *	all early media will passed on to the UA.
  *	pick the one who most recently sent me a criticial threshold of media.
  *	Play only one media stream and ignore others.
  *	The last sdp received override previous one.
  *	First 18x message goes through, rest dropped.
  *	Open voice only with the first one, but answer all of the 18x
  *	We will use the first recieved
  *	I ignore early media
  *	All get relayed - (all rendered leave final choice to recipient UA)
  *	Last early media replaces previous
  *	We update media as the 18x's come in. 200OK media will be the confirmed media channel.
  *	Take the first

Interestingly, 15% of the endpoints supported DHCP option 120.

This is how the endpoints (that actually handled media) described their use of RTCP:
  38% I fully implement RTCP and use the RTCP from my peers
  20% I send some RTCP, and open a port to receive RTCP, but I ignore any packets I receive
  18% I never send RTCP, but I do open a port for receiving (and ignoring) it
  24% I don't even open a port for RTCP

There were 12 (roughly 25%) endpoints testing SRTP support. Keying was predominantly sdes.
Interoperability is not yet high, but more pairs got something working than at SIPit 18.

There were only 4 endpoints supporting comedia.

There were 22 proxies present.

The proxy implementers characterized their handling of infinite loop prevention this way:
   0% I implement loop detection as per the sip-loop-fix draft
  45% I perform RFC3261 loop detection
  45% I don't loop detect, but do pay attention to max-forwards
  10% I don't loop detect or look at max-forwards

I asked proxies "Will you proxy a request with a RURI containing an unknown scheme
(such as splork:) when there is a Route header field whose first value is a SIP URI
you can resolve?" and got these responses:
  32% Yes
  68% No

Half of the proxies in attendence actively participated in session-timer.

There were 9 implementations (41%) that categorized themselves as proxies that would not forward an unknown method.

Two-thirds of the proxies claimed to properly handle SIPS.

None of the proxies made use of 3840 or 3841 information (capabilities and caller-prefs)

There were 19 registrars.

7 of the registrars (37%) accepted non-sip or sips Contacts in a registration
11 (58%) would accept a REGISTER request that had a multipart-mime body (almost all ignored it)
1 would accept an S/MIME signed or encrypted register

Half of the border-elements (B2BUA/SBC-like implementations) could be configured to forward unknown methods.
75% could be configured to forward unknown SDP lines

There were 41 SIP Events implementations present
15 (37%) of them  would send unsolicted notifies (there were 2 more things that ONLY sent unsolicited notifies).

They supported these event packages:
  29 refer
  23 message-summary
  14 presence
  12 dialog
  5  reg
  4  ua-profile (sipping-config-framework)
  4  conference
  2  reg gruu extension (sipping-gruu-reg-event)
  2  certificate/credentials (sip-certs)
  1  session-spec-policy (sipping-policy)
  1  kpml
  0  vx-rtcpxr (sipping-rtcp-summary)

Only 4 (10%) supported winfo

4 supported event-list

37% of the implementation supporting SIP Events supported PUBLISH

Of the 14 implementations supporting event presence, there was support for the following document formats:
  12 base PIDF only
   2 RPID
   0 CIPID
   0 timed presence
   0 PIDF-LO
   0 prescaps-ext

5 implementations supported XCAP
7 supported pres-rules

I asked all the implementations present which P- headers they actively supported:
(I suspect many of the respondents who passively let the headers pass or ignore them answered yes, so these
numbers, more than any others of the above are probably inflated)
  28 P-Asserted-Identity
  21 P-Preferred-Identity
  10 P-Called-Party-ID
  9 P-Associated-URI
  9 P-Access-Network-Info
  8 P-Charging-Vector
  6 P-Visited-Network-ID
  5 P-Charging-Function-Address
  4 P-User-Database
  4 P-DCS-* (any of the P-DCS headers)
  3 P-Media-Authorization