Congestion Control
Connectionless flows
how to define
how to detect
how to respond
how much to overcapacity
is data poisson or fractal?
Congestion control mechanisms I
router-centric
net layer deals with congestion
routers drop packets and notify hosts
host-centric
end systems or processes notice behavior of net and adjust
accordingly
Congestion control mechanisms II
Reservation-based
request type of service from net
admission control
route preservation
Feedback-based
process adjusts to net feedback
explicit by icmp
implicit by behavior of net (packeteer)
Congestion control mechanisms III
window-based
(flow control versus congestion control)
advertised receiver window
rate based
sender requests flow characteristics
Evaluation Criteria
effective resource allocation
increase throughput versus decrease delay
power = throughput / delay
poisson curve of power against load; max point at knee
fairness
equal share of bw
proportinate share of bw according to need
local vs long flows
Queuing Disciplines
FIFO
Fair queuing
Weighted fair queuing
policy mechanisms vs implementation strategies
TCP Congestion Control
Additive Increase/Multiplicative Decrease
Slow-start - exponential increase/decrease
Fast Retransmit and Fast Recovery
Clock problems
Congestion-avoidance Mechanisms
DECbit
RED gateways
implicit host notification by timeout
when and what to drop
MinThreshold and MaxThreshold
MaxP and Weight
partial packet discard
The Next Generation
What else do we want to do with networks?
What are the technical challenges?
What are the tools and technologies?
How it may go from here
What do we want to do next with nets?
Fix some problems
Do applications that require qos
Build interrealm applications
Maintain agency stature
Make new commercial markets
Problems to fix
Routing -
permit policy as a factor
become delay X bw sensitive
Multicast -
discover and join casts
implement multicast routing
IP v6
addresses
mobility
Understand data networks
flows
congestion
Applications that require QoS
Video and visualization
Voice
Bounded ftp's
Synchronized streams for multimedia
Interrealm applications
Interoperable authentication
Common authorization and work flow tools
Metadata standards
Maintain agency status
Esnet
NASA networking
NGI
NSF
Make new commercial markets
videoconferencing
ecommerce and esecurity
next generation networking components
Applications taxonomy
real-time
intolerant
rate-adaptive
non-adaptive but controlled delay
tolerant
adaptive by rate or delay
non-adaptive
elastic
interactive/interactive bulk/ asynchronous
Driver Applications
video and H.320/323
advanced digital libraries and multimedia
multicast services
real-time and delayed collaboration
distributed simulations
environmental crisis management
public information access
remote instrument control
teleimmersion
Comments on Apps and Plumbing
Advanced applications transform high-speed plumbing into value
Advanced plumbing enables advanced applications
Profligate use of bandwidth, per se, does not make an application
'advanced'
Megalomaniac plumbing, per se, does not make the plumbing 'advanced'
Video
broadcast, multicast and pt-pt
digital video - MPEG 1, MPEG 2, DCT, Fractal
lossless compression
lossy compression - spatial and temporal
intraframe and interframe techniques
quantization, run length encoding, motion prediction
Packetized Video Challenges
video is very dependent upon synchronization
difficult transition between a continuous bit stream and a
collection of packets
uncompressed video very high and constant bandwidth, loss tolerant
compressed video much lower and bursty bandwidth, loss intolerant
Real-time and Delayed Collaboration
control and synch of multiple audio and video streams
shared access to information
managed interactions
maintenance of history and audit trails
support of consistency
What are the Technical Challenges?
Laws of Physics
Delay-bandwidth product - Desktops
Quality of service (QOS) - Net hardware
Multicast
Laws of Humanity
Routing and Addressing - Middleware
Applications - Security
Measurements and traffic analysis
Laws of Physics
Delay-bandwidth product - the load in flight
Quality of service (QOS) - differential service requirements for
applications and economics
Desktops - what hardware and software
Network hardware - boxes and lines
Multicast - opens new classes of applications
Delay-Bandwidth product
length of a bit at high speed
effect of long pipes
frames in flight
consequences
buffers, flow control, errors, backplanes, routers
Delay-bandwidth example
Bandwidth Delay x BW
T1 18 KB
Ethernet 122 KB
OC-3 1.8 MB
OC-24 14.8 MB
Tyranny of Latency
assume 100 ms real (48 ms theoretical) RTT cross-country file transfer
assume 1 GBps line, 1 MB to transfer
best result = 0.8 ms +100 ms net thruput 79.4 Mbps
typical 2k packet size, (1,1) sliding window = 0.8 ms + 50 s net
thruput of 160 kbps
Quality of Service (QOS)
Possible QOS Requirements
guaranteed bw
delay/latency
jitter
packet loss
application needs
Possible QOS Technologies
classes of service
overprovisioning
atm
rsvp
diff-serv
Classes of Service
by application
allow certain applications to access high-speed devices
by desktop
highbandwidth to the desktop
multiple parallel campus backbones
WAN PVC's
by IP address (and user)
RED
Traffic shaping
QOS Model Classification
Scope: boundaries of service
intermediate or end-end
Control Model: granularity, duration and locus of control
proxies, levels of aggregation
Transmission parameters
packet loss rate - MTU
bandwidth - delay average and variation
Typical QOS Scenarios
transfer this file within X minutes
provide a response time of less than y ms
guarantee a smooth real-time video stream
support a streaming multimedia presentation
Regardless of approach, QoS Challenges
Issues at the edge
Admission control
Measurement (billing?)
Marking of packets
Issues at the core
Provision of resources to marked packets
Low packet loss in presence of very high delay-bandwidth
products
per-flow knowledge?
How near to complete end-to-end extent is possible
Desktops
equipment - video, codecs
applications
invoking QOS
web as DUI
operating system
buffers and timeouts
context-switching and buffer care
desktops - Memory Bottlenecks
Network hardware
media issues - fiber, LEO's
media signalling issues ATM vs. SONET
router issues - forwarding engine
the fiber shortage ?
Multicast
efficient one-many communications
accomodate dynamic groups
accomodate diversity of connectivity
need at subnet and Internet levels
currently - Mbone; does not scale
"there is no problem that multicast doesn't make harder"
Laws of Humanity
Routing and addressing
Middleware
Applications
Security
Network management
Routing and Addressing
growth of routing tables
non-hierarchical addressing
complexity management
optic routers
Middleware
dynamic location and binding
security
version control
clients and helper apps
Security
Network
Individual
Host and resource
Network Management
measurements and traffic analysis
self-configuration
self-organizing systems
policies and routing
Traffic Analysis
by flow
by application or functional need
sampling techniques
switching makes analysis very hard
Measurement
By end-users for audit
By providers for audit and billing
For fault isolation
For interoperability