Re-thinking the Control Architecture of the...
Transcript of Re-thinking the Control Architecture of the...
Re-thinking the Control Architecture of the Internet
Mark Handley
University College London
Fiber, Cat 5, wireless
Ethernet, WiFi, etc
IP
TCP, UDP, SCTP
SMTP, HTTP, RTP
The Data Plane
The Internet Stack
RTCP
BGP OSPF
MPLS-TE
RSVP RSVP-TE
LDP ARP L2TP
PPP ICMP ICMPv6 IGMP
MLD
IS-IS
DNS
MIP
Shim6 DHCP
SIP SDP NTP SNMP
NetConf
RIP PIM-SM PIM-DM
PWE3
IKE
PCN
RADIUS Diameter
IPsec MPLS
The Control Plane
IGRP
LEAP
MGCP SLP EAP
GRE
GSMP
STUN TURN
ICE
The Internet Control Architecture
Routing
Find best route that complies with policy Congestion Control
Match offered load to available capacity Traffic Engineering
Tune routing policies, establish peerings, provision links, QoS.
Security Access control, authentication
The Internet Control Architecture
Routing
Find best route that complies with policy Congestion Control
Match offered load to available capacity Traffic Engineering
Tune routing policies, establish peerings, provision links, QoS.
Security Access control, authentication
The Internet Control Architecture
Routing
Congestion Control
Traffic Engineering
Medium Timescales
Long Timescales
Short Timescales
The Internet Control Architecture
Routing
Congestion Control
Traffic Engineering
BGP Policies
Deep packet inspection, Traffic shaping
Not much at all
The Internet Control Architecture
Routing
Congestion Control
Traffic Engineering
DDoS Attacks
Congestion Control
The Internet Control Architecture
Routing
Congestion Control
Traffic Engineering
The Internet Control Architecture
Routing
Traffic Engineering
Congestion Control Load
balancing of multihomed sites
The Internet Control Architecture was not a carefully planned separation of functionality, but an accumulation of fixes to specific problems.
Congestion Control (in TCP)
Late addition by Jacobson et al., 1988. Policy Routing (EGP, BGP)
Late addition by Rekhter et al., 1989 onwards.
Traffic Engineering No explicit design prior to mid 1990s Rise of multihoming not foreseen by CIDR, 1993.
Fiber, Cat 5, wireless
Ethernet, WiFi, etc
IP
TCP, UDP, SCTP
SMTP, HTTP, RTP
The Data Plane
The Internet Stack
RTCP
BGP OSPF
MPLS-TE
RSVP RSVP-TE
LDP ARP L2TP
PPP ICMP ICMPv6 IGMP
MLD
IS-IS
DNS
MIP
Shim6 DHCP
SIP SDP NTP SNMP
NetConf
RIP PIM-SM PIM-DM
PWE3
IKE
PCN
RADIUS Diameter
IPsec
The Control Plane
IGRP
LEAP
MGCP SLP EAP
GRE
GSMP
STUN TURN
ICE
INCREMENTAL FIXES
The Power of Legacy
(Pictures courtesy NASA, Darlington Railway Centre)
Network Effect
Metcalfe’s Law: The utility of a
telecommunications network grows with the square of the number of users.
Picture by Derrick Coetzee
Fiber, Cat 5, wireless
Ethernet, WiFi, etc
IP
TCP, UDP, SCTP
SMTP, HTTP, RTP
The Data Plane
The Internet Stack
RTCP
BGP OSPF
MPLS-TE
RSVP RSVP-TE
LDP ARP L2TP
PPP ICMP ICMPv6 IGMP
MLD
IS-IS
DNS
MIP
Shim6 DHCP
SIP SDP NTP SNMP
NetConf
RIP PIM-SM PIM-DM
PWE3
IKE
PCN
RADIUS Diameter
IPsec
The Control Plane
IGRP
LEAP
MGCP SLP EAP
GRE
GSMP
STUN TURN
ICE
INCREMENTAL FIXES Evolution by random mutation and natural selection?
Fiber, Cat 5, wireless
Ethernet, WiFi, etc
IP
TCP, UDP, SCTP
SMTP, HTTP, RTP
The Data Plane
The Internet Stack
RTCP
BGP OSPF
MPLS-TE
RSVP RSVP-TE
LDP ARP L2TP
PPP ICMP ICMPv6 IGMP
MLD
IS-IS
DNS
MIP
Shim6 DHCP
SIP SDP NTP SNMP
NetConf
RIP PIM-SM PIM-DM
PWE3
IKE
PCN
RADIUS Diameter
IPsec
The Control Plane
IGRP
LEAP
MGCP SLP EAP
GRE
GSMP
STUN TURN
ICE
INCREMENTAL FIXES Evolution by random mutation and natural selection?
You are here
You are here
You are here
You are here
You are here
Is it possible to change the Internet architecture in a planned way,
so as to achieve long-term goals?
(or is it only possible to patch the pieces repeatedly until it gets too expensive and unreliable, and eventually something better comes
along and replaces it?)
Spectrum of Networking Research
Cool Hot
Incremental improvement
Blue-sky research
Slow death by a thousand fixes
Undeployable due to network
effects
Possibility to effect planned change
Spectrum of Networking Research
Cool Hot
Incremental improvement
Blue-sky research
Slow death by a thousand fixes
Undeployable due to network
effects
Possibility to effect planned change
Spectrum of Networking Research
Cool Hot
Incremental improvement
Blue-sky research
Slow death by a thousand fixes
Undeployable due to network
effects
Possibility to effect planned change
Spectrum of Networking Research
Cool Hot
Incremental improvement
Blue-sky research
Slow death by a thousand
fixes
Undeployable due to network
effects
Spectrum of Networking Research
Cool Hot
Incremental improvement
Blue-sky research
Slow death by a
thousand fixes
Undeployable due to network
effects
Poss. to
effect planned change
Opportunities to change the game
Routing
Congestion Control
Traffic Engineering
Nearly impossible to replace BGP. Huge network effect.
Nearly impossible to replace TCP. Must co-exist with TCP congestion control, etc
Wish List for Control Mechanisms
Very high robustness, Minimal impact from failures, no downtime. Resists attack.
Load-dependent routing, Move traffic away from congestion Scalable to billions of multihomed mobile nodes
Diverse traffic mix, Even for as yet unforeseen applications and protocols.
Sane economics, Reward network operators for investment.
This talk: opportunities to change the game…
Routing
Congestion Control
Traffic Engineering
The problem of multihoming
Routing
Traffic Engineering
Congestion Control Load
balancing of multihomed sites
Edge links exposed via BGP for fast
recovery
Stresses routing and doesn’t solve the problem well.
Multihoming and Redundancy
Multihoming does provide redundancy.
But routing isn’t a good way to access that redundancy.
Can we access this redundancy better in other ways?
In particular, can we access it at the transport layer, where congestion control can see it?
Smartphone
Mobile client Server
Wifi
3G celltower
Multi-homed server
Client
Server
Load balances between access links
Stripe data from one connection across both paths.
Sending simultaneously across more than one path can provide robustness. ���
If any path dies, TCP can detect it immediately and switch all traffic to the working path.
Client
Server
Sending simultaneously across more than one path can balance load and pool resources.
Each path runs its own congestion control, to detect and respond to the congestion it sees.
But link the congestion control parameters, so as to move traffic away from the more congested paths.
[Kelly & Voice, Key, Massoulie & Towsley]
be less aggressive
be more aggressive
MP-TCP allows a set of disparate links to behave just as if they were a single pooled resource.
The ARPAnet pooled link resources rather than using circuits.
Multipath TCP pools multiple link resources.
ARPAnet resource pooling:
Multipath resource pooling:
Linking the subflows allows a set of disparate links to behave just as if they were a single pooled resource.
This talk: opportunities to change the game…
Routing
Congestion Control
Traffic Engineering
Multipath transport
Economics
Very roughly, the cost of using the Internet is composed of two parts: Access: giving you the possibility to send.
Usage: based on what you send and when you send it.
What is the marginal cost of a packet?
If no-one else wanted to send at that time, there is no cost imposed on anyone from sending a packet.
It only makes sense to charge for usage if that traffic displaces another user’s traffic. Hence we should be charging for congestion.
Common current charging models
Rate-based charging. Eg. charge for 95% percentile rate used.
Volume-based charging. x gigabytes/month volume cap (then some penalty)
These are rough proxies for the real marginal cost: They don’t have the right incentives to cause better user or
application behaviour. They’re inefficient - more efficient players will do better
Example: some applications are latency sensitive, some only care about long-term throughput.
time
b/w
BitTorrent
Windows Update Web page
download
Current incentive model
time
b/w
Incentives when charging for congestion volume
Charging for congestion volume encourages machine-to-machine traffic to move to uncongested times.
Users probably won’t change their behaviour.
Want to watch TV at peak time.
But a lot of apps are flexible in when they send.
And even the interactive ones may be able to get creative with prefetching, etc.
The missing mechanism
ISPs can’t charge for congestion, because they can’t see it properly. End-system can see it.
Router that drops packets has some idea.
No-one else can see the big picture, so there’s no framework to pass on the costs of congestion.
Congestion Exposure���(general term for Bob Briscoe’s re-feedback, re-ECN, etc)
Indicate to an ISP the congestion seen by traffic downstream of them.
Congested Links
Feedback: x packets saw congestion
Downstream congestion Upstream congestion
Congestion Exposure is an enabler for sane economics.
Allows everyone forwarding packets to see the cost those packets will impose on the downstream networks. Enables a network-neutral response that differentiates
traffic based on its real cost. Encourages well-behaved applications. Needs to be some pass-on of the cost to customers for
them to prefer well-behaved applications. Lots of potential for innovation here.
Opportunities to change the game…
Routing
Congestion Control
Traffic Engineering
Multipath transport
Congestion exposure
Changing the Game: Potential Outcomes
Multipath transport moves traffic away from congested paths onto uncongested paths.
Congestion exposure encourages well-behaved applications that move traffic from congested times to uncongested times.
Between the two, the network can be much more robust, cope with much wider ranges of load, and better handle the demanding applications of tomorrow.
Re-thinking the Internet Control Architecture.
Routing
Congestion Control
Traffic Engineering
By taking a holistic view of control, we can enable whole new degrees of freedom.
By working with existing protocols we can do this in a way that stands some chance of being deployed.
Spectrum of Networking Research
Cool Hot
Incremental improvement
Blue-sky research
Slow death by a thousand fixes
Undeployable due to network
effects
Possibility to effect planned change
Don’t neglect phase two of
your research!
Thanks
Especially to: Damon Wischik Costin Raiciu Bob Briscoe Phil Eardley
Lars Eggert Sebastian Barre Alan Ford Marcelo Bagnulo
Multipath TCP: http://nrg.cs.ucl.ac.uk/mptcp/