Congestion Control

Download Report

Transcript Congestion Control

Congestion Control
Andreas Pitsillides
University of Cyprus
1
Congestion control problem
growing demand of computer usage requires:

efficient ways of managing network traffic to avoid or limit
congestion in cases where increases in bandwidth not
desirable or possible.
generally accepted that network congestion control
problem remains critical issue and high priority,

especially given growing size, demand, and speed
(bandwidth) of increasingly integrated services network.
One could argue that

network congestion unlikely to disappear in near future.
Furthermore congestion may become
unmanageable
 unless effective, robust, and efficient methods
for congestion control are developed.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
2
Current scene
despite vast research efforts, still no universally
acceptable solutions:
 control
solutions for TCP transported traffic
increasingly becoming ineffective,
 cannot easily scale up even with:
 “fixes” (improved round trip time
measurement, Slow-start and congestion
avoidance, Fast retransmit, fast recovery
algorithms, Improved congestion indication
using delay (rather than loss) as feedback.
 new approaches (RED, ECN, MPLS)
 new architectures (diffserv, intserv,)

Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
3
Current scene (cont.)
 non-TCP applications



As demand for streaming applications increases,
important to ensure can co-exist with current TCP
streaming media should be subjected to similar rate
controls as TCP traffic
newly developed (also largely ad-hock) strategies are
also not proven to be robust and effective

examples include model based and equation based
approaches.
 Even though based on a model, model is not dynamic,
derived control strategy is ad-hock and not proven
with regard to its properties.
 Asynchronous Transfer Mode (ATM)

also witnessed similar approach, with performance of
vast majority of congestion control schemes proposed for
solution of Available Bit Rate (ABR) problem not proven
analytically.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
4
Why problem still not
solved?
 In part, due to lack of structured approach, and
 lack of strong theoretical foundation in
stabilising controlled systems,

Most proposed schemes are developed using
intuition and simple (ad-hock) non-linear designs.



Using simulation, these simple schemes demonstrated to
be robust in variety of scenarios.
problem is that very little known why these methods work
and very little explanation can be given when they fail.
Since designed with significant non-linearities,
based mostly on intuition (e.g. two-phase—slow start and
congestion avoidance—dynamic windows, binary feedback, …)
 analysis
of closed loop behaviour difficult, if at all possible,
even for single control loop networks.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
5
Why problem still not solved? (cont.)
interaction of additional non-linear feedback loops
can produce unexpected and erratic behaviour.


Empirical evidence demonstrates poor performance and
cyclic behaviour of the controlled TCP/IP Internet (also
confirmed analytically).
becomes worse



as link speed increases (hence bandwidth-delay product, and
thus feedback delay, increases)
as demand on network for better quality of service increases.
for WAN networks


multifractal behaviour has been observed,
suggested that this behaviour—cascade effect—may be related
to existing network controls.
 Clearly, more effective congestion control schemes
are needed to prevent serious economic losses and
possible "meltdown" of the Internet.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
6
Two examples of existing
disciplines with strong
theoretical foundation
control systems theory



rich experience in controlling complex systems,
often concentrating (due to the difficulty) on single control
loops to stabilise the whole system (by assuming if locally
stable, then also globally—some theoretical foundation
exists).
traditionally linearising model to apply linear control systems
theory  new results in non-linear theory allow application
Pricing theory



has proven useful for stabilising complex interactions in
human centred systems,
aiming to balance supply and demand.
Usually distributed algorithms, which through successive
iterations reach stability
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
7
IDCC: an example (with Petros
Ioannou and L. Rossides)
Starting with a simple dynamic fluid flow
model:

developed using packet flow conservation considerations and
by matching the queue behaviour at equilibrium
Design a non-linear adaptive robust controller
(IDCC - integrated dynamic congestion controller)


a specific problem formulation for handling multiple
differentiated classes of traffic, operating at each output port
of a switch is illustrated.
following same spirit adopted by IETF Diff-Serv for Internet
define three classes of aggregated behaviour.

Premium, Ordinary, and Best Effort Traffic Services.
analytical performance bounds derived, for
provable controlled network behaviour.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
8
Control concept
r (t)
Integrated Dynamic Congestion
Controller
(IDCC)
Allowed common rate
sent to the Ordinary
Traffic sources
Ordinary
rin ( t )
Traffic
b (t )
Best effort
traffic
Andreas Pitsillides,
University of Cyprus
xr (t )
x p (t )
Premium
 (t )
Traffic p
Incoming
traffic
xref
(t)
p
xrref (t)
C p (t )
references
C p (t )
Fixed service
rate Cserver (e.g.
155 Mb/s)
Cr (t)
Scheduler
with server
buffer
Instantaneous
left-over capacity
Cost 257 final seminar
27-29 September 2000
9
Dynamic model
For a packet buffer:
x (t )   f out (t )  f in (t )
For M/M/1 queue
x (t )
x (t )  
C (t )   ( t )
1  x (t )
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
10
Simulative comparison
700
x(t),
queue
length 600
OPNET simulation
500
400
300
200
100
0
fluid flow model solution
-100
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
time (s)
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
11
Another dynamic fluid
flow model
for TCP window:
W (t ) 
1
W (t )W (t  R(t ))

p(t  R(t ))
R(t )
2R(t )
N (t )
x(t ) 
W (t )  Cserver
R(t )
Andreas Pitsillides,
University of Cyprus
x(t )
R(t ) 
 Tp
Cserver
Cost 257 final seminar
27-29 September 2000
12
Developed Control
strategy
Premium Traffic Service (eq. 1, 2, 3)



1  x p (t )


 p x p (t )  k p (t )  
C p (t )  max 0, min Cserver ,  p (t )
x p (t )






0
if x p (t )  0.01

 p (t )  1.01x p (t )  0.01 if 0.01  x p (t )  1

1
if x p (t )  1

k p (t )  Pr  p x p (t )
Ordinary Traffic Service (eq. 4)



xr ( t )
r (t )  max 0, min Cr (t ), Cr (t )
  r xr ( t )  
1  xr ( t )



Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
13
Theoretical evaluation
A1. Proof of stability of Premium Traffic control
strategy
Theorem A1. The control strategy described
by the equations (1-3) guarantees that



queue length is bounded
allocated Capacity<=Server Capacity
queue length converges close to the reference
value with time, with an error that depends on the
rate of change of the traffic input rate.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
14
Theoretical evaluation (cont.)
A2. Proof of stability of the Ordinary Traffic
control strategy
Theorem A2. The control strategy given
by equation (4) guarantees that
queue length is bounded.
 When bandwidth becomes available the
queue length approaches the reference
value with time.

Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
15
Simulative evaluation
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
16
Steady state and
transient behavior
Qureue length
Ref=100
ref=100
ref-=50
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
Switch 2 time evolution of
Premium Traffic queue
length for a LAN and WAN
for 140% load demand.
Note that as feedback
information is local, there
is no deterioration in
performance due to
increased WAN
propagation delay.
17
Steady state and transient behavior
(cont.)
Ref=900
Ref=600
Ref=300
Switch 2 time evolution of Ordinary Traffic queue length for (a) a LAN and
(b) WAN for 140% load demand. (control period varies between 32
celltimes0.085 msec to 353 celltimes0.94 msec)
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
18
Steady state and transient behavior
(cont.)
Typical behaviour of the time evolution of the common calculated
allowed cell rate at Switch 2 for (a) LAN and (b) WAN.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
19
Steady state and transient behavior
(cont.)
Typical behavior of time evolution of transmission rate of controlled
sources using Switch 2 for (a) LAN and (b) WAN configurations.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
20
Network test configuration
for demonstrating fairness
3-hop traffic start transmitting at t=0
the one 1-hop-a traffic at switch 0 is next started at t=0.2
the two 1-hop-b sources atswitch 1 are started at t=0.4
the three 1-hop-c sources are started at t=0.6
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
21
fairness - LAN
3-hop
1 hop-a
1 hop-a
3-hop
1 hop-a
3-hop
1 hop-b
1 hop-b
3-hop
1 hop-c
Allocation of bandwidth to Ordinary Sources for LAN. All
sources dynamically allocated their fair share at all times.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
22
fairness - WAN
3 hop
1 hop-a
1 hop-a
3 hop
1 hop-a
3 hop
1 hop-b
1 hop-b
3 hop
1 hop-c
Allocation of bandwidth to Ordinary Sources for WAN. All
sources dynamically allocated their fair share at all times
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
23
fairness - WAN
Allocation of bandwidth
to the Ordinary Sources
at Switch 2. Observe
that the top 3 figures are
for local sources and the
last one is for a 3 hop
source located about
12000 kms away from
the switch.
All sources are allocated
their fair share
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
24
Behaviour of control
Insensitivity of control to the value of the
control update period

32 celltimes0.085 msec to 353 celltimes1 msec
Robustness of control design constant to
changing network conditions

for diverse traffic demands ranging from 50%140% and source location (feedback delays) up to
about 250 msec RTT, as well control periods
ranging from 0.085 msec to 1 msec. For all
simulations the behaviour of the network remains
very well controlled, without any unacceptable
degradation
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
25
IDCC properties
provable stable and robust behaviour at each port,

and by tightly controlling each output port,
performance expected to be tightly controlled.
overall network
high utilisation with bounded delay and loss performance
good steady state behaviour, with no observable oscillations
good transient behaviour, i.e. fast rise and quick settling times
Uses minimal information to control system and avoids
additional measurements and noisy estimates:





Uses only one primary measure, namely queue length
Does not require per connection state information, queuing, or
servicing at the switch
Does not require any state information about set of connections
bottlenecked elsewhere in network (not even count)
Computes Common Ordinary Traffic allowable transmission rate only
once every Ts msec (control update period) thereby reducing
processing overhead.
controller fairly insensitive to value of Ts.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
26
IDCC properties (cont.)
Achieves max/min fairness in a natural way without additional
computation or information
can guarantee minimum agreeable service rate without additional
computation
works over wide range of network conditions, such as RTT
(feedback) delays, traffic patterns, and controller control intervals,
without change in control parameters
works in integrated way with different services (e.g. Premium Traffic,
Ordinary Traffic, Best Effort Traffic) without need for any explicit
information about their traffic behaviour
proposed control methodology and its performance is independent of
size of queue reference values.

network operator can be more or less aggressive and steer performance, in
accordance with current network and user needs, using global consideration.
Has simple implementation and low computational overhead
features very small set of design constants,

can be easily tuned from simple understanding of system behaviour
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
27
Conclusions for IDCC
generic scheme for congestion control.





uses integrated dynamic congestion control
approach (IDCC).
specific problem formulation for handling multiple
differentiated classes of traffic, operating at each
output port of a switch illustrated.
derived from non-linear control theory using a fluid
flow model.
analytical performance bounds derived, for
provable controlled network behaviour.
divide traffic into three basic types of service, in
same spirit as those adopted for Internet Diff-Serv
i.e. Premium, Ordinary, and Best Effort.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
28
Conclusions for IDCC (cont.)
As shown earlier, proposed control algorithm
possesses a number of important attributes
works in integrated way with different services
has simple implementation and low
computational overhead,
features a very small set of design constants
that can be easily set (tuned) from simple
understanding of system behaviour.
 These attributes make proposed control
algorithm appealing for implementation in
real, large-scale heterogeneous networks
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
29
further work for IDCC
In this paper full explicit feedback was used in
the simulations, signalled using RM cells in
an ATM setting.
challenging task is to investigate other explicit
(e.g. single bit feedback as in ECN proposal
for IP) and implicit (end-to-end) feedback and
signalling schemes.

A comparative analytic and simulative evaluation
between the different feedback and signalling
schemes is a topic for future research.
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
30
General Recommendations
 Advocate a structured and formal approach to
designing congestion control systems


could be from other fields with solid theoretical foundation,
possibly drawn from stabilising (controlling) large scale,
complex systems
encourage collaboration with other disciplines
 Integrate with other control functions and study their
interactions (e.e. with routing and CAC)
 A common simulative framework (CSF) and pilot test
bed environment (e.g. ns 2 could be such a
simulative test-bed)

with well known and understood scenaria that test the
properties of proposed algorithms

e.g. dynamic properties, robustness, large scale deployment
aspects, steady state behaviour, and so on…
Andreas Pitsillides,
University of Cyprus
Cost 257 final seminar
27-29 September 2000
31