[SIP SIMPLE client] MSRP might just not be the right choice for tunneling RFB

Denis Bilenko denis at ag-projects.com
Tue Jan 27 14:04:49 CET 2009


Willem Toorop wrote:
> Hi Adrian, Michiel, Ruud, Denis et al.
>
> After I putted the pieces together as described earlier, I kept
> getting this error:
>
> Traceback (most recent call last):
>   File "./21_sip_offer_desktop", line 85, in <module>
>     main()
> ...
>   File
> "/usr/lib/python2.5/site-packages/eventlet-0.9pre-py2.5.egg/eventlet/coros.py",
> line 110, in reset
>     assert self._result is not NOT_USED, 'Trying to re-reset() a fresh
> event.'
> AssertionError: Trying to re-reset() a fresh event.
>
I hope this is not happening anymore. If you do have this or other weird
error please let me know how to reproduce them.

> Looking at twisted/internet/abstract.py I saw that everything written
> is buffered and that the producer is paused when the buffer is full.
> The producer is in our case the process reading from the vncserver and
> writing to the msrp transport. Only nothing was actually paused. The
> send_message method, always returns immediatly.
> I've tried several things (like testing for the buffersize, and then
> waiting for a while until the buffer would shrink again), but I finaly
> concluded that
>
> The write method of class MSRPTransport in file msrplib/transport.py
>
>     def write(self, data):
>         if self.traffic_logger:
>             self.traffic_logger.report_out(data, self.transport)
>         return self.transport.write(data)
>
> should be altered to:
>
>     def write(self, data):
>         if self.traffic_logger:
>             self.traffic_logger.report_out(data, self.transport)
>         GreenTransportBase.write(self, data)
>
> as the baseclass GreenTransportBase, *does* implement a mechanism in
> which it can be paused by twisted calling the pauseProducing function.
great observation, fixed. Thanks!

> After making the proposed alteration, the vncserver and client
> connected succesfully, but after a while, garbage is written the the
> viewer. In the vncserver log it mentions the creation of a xdamage
> object, and shortly afterward mentions the resetting of the connection
> by the client.
> I suspect that the messages from the vncserver to the client might not
> arrive in the right order. I think the MSRP-relay just tries to sent
> out all messages it receives as quickly as possible. A short messages
> could arrive earlier then a big messages, even if the short message
> was sent after the big message. I should investigate this further to
> be sure... but how?.
> Increasing the message size of the server does work (maybe because all
> the related RFB messages then are more likely to stick together), but
> the performance in the viewer is then extremely poor. Screen updates
> take tens of seconds.
>
> When using MSRPTransport.deliver_message in stead of send_message, the
> performance is much better! Small updates take only 3 seconds or less,
> but big update still take tens of seconds. I'm under the impression
> that smaller message sizes perform better with deliver_message then
> big message sizes, but it doesn't differ much. It is still unusably slow.
> I understand (from reading the source) that deliver_message only
> checks the first hop, and doesn't actually guarantee that the message
> are actually delivered. Perhaps waiting for this answer just slows
> down the process so much, so that messages are able to arrive in the
> right order.
>
> It is also interesting to see how x11vnc adjust its estimated link
> rate. Only with deliver_message it estimates it to slower than LAN
> speed. Maybe x11vnc adjusts its update alogrithm to the link speed.
>
> I think ideally with RFB over MSRP we would like to use send_message,
> in stead of deliver_message. We do not need to be confirmed of the
> message being deliverd at the MSRP-Relay. We do however like te be
> able to process the next message only as the previous one is sent out
> over the tcp connection. Even with the altered write method,
> send_message still doesn't allow this. The transport object (from
> twisted) seems to use a buffering mechanism with a seperate threat
> actually sending out all the messages.
>
> I understand that the usage of those twisted classes will give better
> performance at an instant-messaging-user-interface. The control is
> directly returned to the process displaying and inputting messages,
> without waiting for the old message being sent out, but for a regular
> tcp session this is not ideal.
>
> I see another problem with TCP sessions over MSRP in general. With
> regular TCP an answer to a message can be returned with the
> confirmation of the reception of the message. With MSRP this is not
> possible. MSRP acts as two unidirectional streams which are guaranteed
> to be delivered, but the connection between the reception of a message
> and the response is not there anymore. This is no problem with streams
> that have a 'command, reply' structure (such as smtp, pop, http etc),
> but it is a problem when there is a stream of messages and a answer
> should be able to be sent back with the acknowlegment of a specific
> message from that stream. Such a message, when tunneling over MSRP,
> could be received at a much later time.
> I'm not under the impression that this is a problem with the RFB
> protocol, but I can imagine that a better update mechanism could be
> chosen by the vncserver when it has a realistic idea of how quickly
> its messages are received by the client. At the moment it seems to
> estimate this on the basis of the time it takes to actually sent out a
> message (which we can't because of twisted buffering). An
> implementation using a TURN server would also have this problem: The
> vncserver would then estimate its speed as the speed it takes for the
> TURN server to receive the messages.

There is indeed a buffering in twisted's transport that's inherited by
MSRPTransport.
It really is unnecessary and makes a lot of stuff harder to implement. I
plan to get rid of it
in the nearest future. Once that is done, send_message would block until
the data hits
OS' "send" buffer.

The next thing on the roadmap for msrplib is to make outgoing responses have
higher priority than regular data, so if there's a response in the
queue, the chunk
that is currently being sent will be interrupted.

Without such mechanism deliver_message, which waits for the response,
may indeed
take a rather long time to return.

BTW, the newer chat script has --msrp-tcp option to establish TCP
connection instead of TLS.
Combine that with -r none and you can cut the relay out of the equation
to see if it makes any
difference. (Establishing direct TLS connections is not possible at the
moment)






More information about the SIPBeyondVoIP mailing list