110 likes | 202 Vues
TCP, ALF and RM Futures. Positions for the ICNP 98 Future Transport Panel Allison Mankin mankin@east.isi.edu http://www.east.isi.edu/RMRG. Ineffability of TCP. The Internet had to have byte-stream service: imagine building any large net without assured file transfers
E N D
TCP, ALF and RM Futures Positions for the ICNP 98 Future Transport Panel Allison Mankin mankin@east.isi.edu http://www.east.isi.edu/RMRG
Ineffability of TCP • The Internet had to have byte-stream service: imagine building any large net without assured file transfers • What other transports must be everywhere? • Perhaps none - depending on your answer to whether UDP is a transport
Immobility of TCP • Improving TCP is frequently envisioned, but may now be too hard to effect • TCP Large Windows case - it’s out there, but offered window field adaptation piece never happened - need for it was seen when Internet had just grown to 100,000 hosts • Contrast RTP deployment. With hooks for Application Layer Framing, protocol fits many niches, and published transports using RTP alone number 10’s per year
Invisibility of ALF Transports • Many new transports on the net are application-framed (ALF). Specs may not be public. Example: Real Network tm RTP/RTCP UDP or TCP IP or IP Multicast
Inevitability of Reliable Multicast Transports • RM, such as: SRM (shared interactive Pressure to have them, especially for software update distribution (latest pull seen in Gates’ Sep 98 MegaServer memo) • Megaserver is described as using a bulk transfer RM; many other types of multicast, such as SRM (shared interactive media), MTP II (image-oriented) • Multicast transports do not yet have congestion control (but they should, see also RFC 2357)
Intractability of Reliable Multicast Transports • Despite demand, none yet deployed on general Internet • Hold-ups from congestion problems (that break the product performance), reported even on the dedicated nets • Intractable too strong - simple congestion control for bulk transfer RM is on track in a design team of the RM Research Group
Improvability of Reliable Multicast Transports • End2End Research Group: “It’s not an Internet transport protocol if it does not adapt to congestion” • Reliable Multicast Research Group: designing simplest possible RMCC • Enabled by progress on the “TCP-friendly” equations
Inability to End This Joke? Some title words left (iniquities…) but... A NAAH
Multicast Congestion Control Research Has Been Delayed, Besides Being Hard B A [R]M flow TCP flow A fallacy: Congestion on the way to B has been “paid for” because one flow went to A’s three children instead of 3.
Aspects of TCP-Friendly CC [R]M flow TCP flow TCP sees RM adapting to its different bottlenecks. Potential oscillations. What share for TCP versus what share for a multicast flow is orthogonal to stabilizing both their CC behaviors
Three Positions Taken • TCP will remain pretty much as is • ALF transports will continue filling niches. Traffic researchers need to be observing them • TCP-Friendly congestion control will dominate the largest, toughest class of Future Transports, the RM’s