1. Peter Parnes, CDT1/22 Media Scaling of IP-Multicast Streams in Heterogeneous Networks Peter Parnes LTU-CDT/Marratech Roxy Workshop 980921-23 Media Scaling of IP-Multicast Streams in Heterogeneous Networks Peter Parnes LTU-CDT/Marratech Roxy Workshop 980921-23

2. Peter Parnes, CDT2/22 Overview  Background  Current Problem  Proposed Solutions  mStar  Current Status in the Internet  Summary  Background  Current Problem  Proposed Solutions  mStar  Current Status in the Internet  Summary

3. Peter Parnes, CDT3/22 Background  “Broadcasts” of real-time media on the Internet is becoming more and more important.  It is very central to the Roxy project.  If the used system shall scale, IP- multicast HAS to be used!  “Broadcasts” of real-time media on the Internet is becoming more and more important.  It is very central to the Roxy project.  If the used system shall scale, IP- multicast HAS to be used!

4. Peter Parnes, CDT4/22 Requirements and Restrictions  Best-effort delivery  Reliability not required  Applications have to be adaptive, i.e. have to adapt to network congestion and be able to handle different configurations.  Best-effort delivery  Reliability not required  Applications have to be adaptive, i.e. have to adapt to network congestion and be able to handle different configurations.

5. Peter Parnes, CDT5/22  Which bandwidth should be used when transmitting a real-time media stream over heterogeneous networks? Internet Sender 100Mbps Local Receiver 100Mbps Internet Receiver 500Kbps ISDN Receiver 128Kbps

6. Peter Parnes, CDT6/22 Proposed solutions  Max/Min client bandwidth  Simulcast  Network transcoders  Receiver driven Layered Multicast - RLM  Bandwidth Guessing - TCP friendly  Active Networks  Active Services  Max/Min client bandwidth  Simulcast  Network transcoders  Receiver driven Layered Multicast - RLM  Bandwidth Guessing - TCP friendly  Active Networks  Active Services

7. Peter Parnes, CDT7/22 Max/Min client bandwidth  Just ignore some set of receivers  Send the stream with high bandwidth  Ignore low bandwidth receivers  Send the stream with low bandwidth  Force high bandwidth receivers to use low quality  Does not take congestion into account  Just ignore some set of receivers  Send the stream with high bandwidth  Ignore low bandwidth receivers  Send the stream with low bandwidth  Force high bandwidth receivers to use low quality  Does not take congestion into account

8. Peter Parnes, CDT8/22 Simulcast  Send the same stream with different encodings from the sender and let the receivers choose what they want to receive.  Can be very expensive CPU wise  Wastes bandwidth on shared links  Does not take congestion into account in the way it is being used today.  Used in mStar (more later)  Send the same stream with different encodings from the sender and let the receivers choose what they want to receive.  Can be very expensive CPU wise  Wastes bandwidth on shared links  Does not take congestion into account in the way it is being used today.  Used in mStar (more later)

9. Peter Parnes, CDT9/22 Network Transcoders  A common approach is to deploy transcoders on the boundaries between different networks.  Transcoding, mixing, downscaling  E.g. transcode MJPEG to H.261 when the traffic leaves a campus (high bandwidth network).  E.g. mStar mTunnel can transcode when tunnelling mcast data (more later).  A common approach is to deploy transcoders on the boundaries between different networks.  Transcoding, mixing, downscaling  E.g. transcode MJPEG to H.261 when the traffic leaves a campus (high bandwidth network).  E.g. mStar mTunnel can transcode when tunnelling mcast data (more later).

10. Peter Parnes, CDT10/22 Receiver driven Layered Multicast - RLM  Divide the stream into a hierarchy of exclusive additive layers  Each layer is multicasted to a different group loop: if no_congestion then join next group to get higher layer else leave group to drop highest layer  Divide the stream into a hierarchy of exclusive additive layers  Each layer is multicasted to a different group loop: if no_congestion then join next group to get higher layer else leave group to drop highest layer

11. Peter Parnes, CDT11/22 RLM Problems  How to detect congestion caused by my tests or by others  Shared learning proposed  Does it scale?  Today long timeout in mcast forwarding trees  Might lead to false interpretation of the current situation  Is not “nice” to TCP  How to detect congestion caused by my tests or by others  Shared learning proposed  Does it scale?  Today long timeout in mcast forwarding trees  Might lead to false interpretation of the current situation  Is not “nice” to TCP

12. Peter Parnes, CDT12/22 Bandwidth Guessing  In early 97 a proposal called “TCP- Friendly” was distributed.  Describes a way of estimating the bandwidth between a sender and a receiver based on RTT and current packet drop.  Takes TCP into account and will be a “nice” participant in the network  In early 97 a proposal called “TCP- Friendly” was distributed.  Describes a way of estimating the bandwidth between a sender and a receiver based on RTT and current packet drop.  Takes TCP into account and will be a “nice” participant in the network

13. Peter Parnes, CDT13/22 BW Guessing Problems  Hard to calculate RTT accurately  Works only for “broadcast” situations  Not very tested yet  Hard to calculate RTT accurately  Works only for “broadcast” situations  Not very tested yet

14. Peter Parnes, CDT14/22 Active Networks  The latest “buzzword” network research topic (since ATM is practically dead)  Basic idea:  Allow injection of small programs into network nodes  Network nodes perform computations on user data  The latest “buzzword” network research topic (since ATM is practically dead)  Basic idea:  Allow injection of small programs into network nodes  Network nodes perform computations on user data

15. Peter Parnes, CDT15/22 Active Networks...  Two Different Approaches  Code and control is handled out-of-band  Each packet carries miniature programs (capsules)  Allows networks to be modified “on- demand”  Opens a completely new area for real- time media scaling  Two Different Approaches  Code and control is handled out-of-band  Each packet carries miniature programs (capsules)  Allows networks to be modified “on- demand”  Opens a completely new area for real- time media scaling

16. Peter Parnes, CDT16/22 Active Networks...  Issues:  Safety, fairness, appropriate architecture, common programming model, robustness  Status:  At the very beginning  A very “political” problem “I dare You to run code in my router!” Steve Deering - Cisco (currently :-)  Issues:  Safety, fairness, appropriate architecture, common programming model, robustness  Status:  At the very beginning  A very “political” problem “I dare You to run code in my router!” Steve Deering - Cisco (currently :-)

17. Peter Parnes, CDT17/22 Active Services  Deploy user controllable programs-pads in the network.  Users can deploy their own transcoding programs and can easily up-grade these when needed  A system for this is currently being deployed and tested on Berkeley Campus  Deploy user controllable programs-pads in the network.  Users can deploy their own transcoding programs and can easily up-grade these when needed  A system for this is currently being deployed and tested on Berkeley Campus

18. Peter Parnes, CDT18/22 mStar  mTunnel contains support for media- aware rescaling, transcoding, mixing and switching of audio and video.  Could be easily be extended for general transcoding between different mcast groups.  Simulcast is currently being used in mStar Pro for the electronic corridor.  BW Guessing and RLM based approaches should be further investigated.  mTunnel contains support for media- aware rescaling, transcoding, mixing and switching of audio and video.  Could be easily be extended for general transcoding between different mcast groups.  Simulcast is currently being used in mStar Pro for the electronic corridor.  BW Guessing and RLM based approaches should be further investigated.

19. Peter Parnes, CDT19/22 Status in Internet  Almost all traffic is still sent using Unicast - transcoding at the server  Network transcoders probably most common  Simulcast less common than one could imagine (lack of good support in today’s applications).  Almost all traffic is still sent using Unicast - transcoding at the server  Network transcoders probably most common  Simulcast less common than one could imagine (lack of good support in today’s applications).

20. Peter Parnes, CDT20/22 Summary  A number of more or less proposed solutions:  Max/Min client bandwidth  Simulcast  Network transcoders  Receiver driven Layered Multicast - RLM  Bandwidth Guessing - TCP friendly  Active Networks  Active Services  Still a lot of research needed  A number of more or less proposed solutions:  Max/Min client bandwidth  Simulcast  Network transcoders  Receiver driven Layered Multicast - RLM  Bandwidth Guessing - TCP friendly  Active Networks  Active Services  Still a lot of research needed

21. Peter Parnes, CDT21/22 Questions? peppar@cdt.luth.se http://www.cdt.luth.se/~peppar/ http://www.cdt.luth.se/mStar/ http://www.marratech.com/ peppar@cdt.luth.se http://www.cdt.luth.se/~peppar/ http://www.cdt.luth.se/mStar/ http://www.marratech.com/

22. Peter Parnes, CDT22/22 Multicast vs. Unicast