Transcript slides

Panel Contribution:
The Future of Transport—
Evolution or Revolution
Steve Deering
[email protected]
Above My Layer!
• transport protocol development is crucial to
restore the thin waist of the hourglass
– an architecturally clean alternative to various “transport
helper” hacks inside the network
– datalink protocol development also important
• multicast transport development needed to
remove obstacle to IP multicast deployment
• will become even more essential if end-toend encryption (IPsec) becomes common
Better IP Layer Support
for Transport Protocols?
• disincentives for bad (non-cooperative)
behavior?
– e.g. “penalty boxes”
– definition/characterization of bad behavior
• better fairness?
– e.g., stochastic fair queueing
• better congestion feedback?
– e.g., ECN
Challenging Link/Path Properties
(Vern Paxson’s list)
• long delay—affects timer estimation,
congestion adaptation
• high bandwidth-delay product—difficult to
fully utilize and still respond to congestion
• variable bandwidth—difficult to track
throughput target
• varying delay—may affect accuracy of RTO
estimation
Challenging Link/Path Properties
(cont.)
• links with link-layer flow control—may
interact adversly with end-to-end control
• high error rates—performance effects when
interpreted as congestion
• inconsistent error rates—can defeat
attempts to tell loss from congestion
• expensive connection set-up or retention—
hostile to soft-state protocols
Challenging Link/Path Properties
(cont.)
• very low bandwidth—creates pressure for
blurring layers and custom demuxing
• unidirectional—breaks most current IP
routing (not really a transport issue)
• non-transitive reachability—breaks subnet
assumptions (not really a transport issue)
• large clouds—(not really a transport issue)
Challenging Link/Path Properties
(cont.)
• paths that reorder packets—defeats
common optimizations, hueristics
• multipathing—multiple path properties to
estimate
• intermittent outages—may excessively
stretch RTT estimates
• very small MTU—may create black holes if
MTU discovery doesn’t work
Challenging Link/Path Properties
(cont.)
Question is, which of these can/should be:
• handled at transport layer?
• handled at datalink layer?
• ignored?
One Pragmatic Suggestion
• use ALF (or at least above-UDP)
approaches to enable easier deployment