Core Stateless Fair Queueing

1
Core Stateless Fair Queueing

• Ion Stoica Hui Zhang Scott Shenker
• CMU CMU Xerox PARC

2
Document Organization

• Provide a brief description and overview of
  Coreless Stateless Fair Queing
• Describe the algorithms supporting the model
• Simulations and comparisons with differnet loads
• Conclusion and results

3
Introduction

• The purpose of this paper is to achieve
  congestion control by allocation of bandwidth in
  a fair manner.
• The architecture proposed is a set of algorithms
  which are simple and inexpensive to implement on
  a contiguous portion of network with well defined
  interior and edges.

4
Mechanisms for Congestion Control and Fairness

• FIFO Drop tail strategy.
• RED Buffer management, probabilistically drop
  packets to keep the occupancy between
  the max and min thresholds
• FRED(Flow Random early drop) Extends RED
  to achieve fairness.
• DRR(Deficit Round Robin) Employs
  sophisticated weighted fair queuing.

5
Architecture for Core Stateless Fair Queuing

• Edge Router
• a) Maintains Flow per state.
• b) Label the packet with the rate estimate
• Core Router
• a) Uses FIFO Queuing
• b) Probabilistic dropping
• depending on label, fair rate of router
  depending on aggregate traffic

6
Functions of Edge and Core Routers

• Classify the packets to a flow
• Label the packet
• Update the fair share rate estimation for the
  outgoing link
• Update the flow rate estimation
• Forward the packet depending on dropping
  probability max(0,1-a/r(i))
• Relabel the packet.

7
Fluid model and Packet Algorithm at the core

• Suppose each flow arrival rate ri(t) is known
• If its fair share rate is a(t), then the service
  received by the flow at time t will be min(ri(t)
  ,a(t))
• Thus the total outgoing traffic from the router
  would be the ? (min(ri(t) ,a(t)) C
• The fraction of packet dropped from each flow is
  calculated by ((ri(t) - a(t))/ ri(t).

8
Flow Arrival rate at the edge router

• Use a exponential averaging system.
• The estimated rate of arrival of each new packet
  is computed from the inter arrival time, the
  length of the packet and the previous rate which
  is thus updated
• Rnew (1-e-Ti/K)Li/Ti e-Ti/KRold
• where Li is the length of the packet at ti
• and T is the time between packet ti-1 and ti

9
Link Fair Rate Estimator

• The estimated rate is put into the label by the
  edge router.
• At the core router the fair rate is estimated
  from the algorithm of aggregated arrival rate and
  aggregated accepted rate which are updated from
  each arrival of packets.

10

• Link congested
• F(a)?min(ri(t), a(t))
• a(t) is the fair share rate
• F is the estimated fair packet accepance rate of
  all the traffic.
• Agg(n )(1-e-T/K)L/T e-T/KAgg(o)
• Agg(n)The estimated aggregate arrival rate
• F is calculated similarly
• The estimate for fair share rate is updated like
• anew aold C/F

• Link not congested
• a is said to the largest rate of any active flow
  r(i) and aggregate received traffic
• Agg(n) ltC during an interval of length Kc

11
Algorithm

• The edge router estimates the rate using
• Rnew (1-e-Ti/K)Li/Ti e-Ti/KRold
• and puts in the label
• The core router finds the probability of
  dropping the packet using the fair rate estimator
  a
• Drop Probmax(0,1- a /p.label)
• The rate estimator a is updated with each packet

12
Label rewriting

• If there is congestion at core router, the packet
  labels need to be updated at the core router
  which is done by
• Lnew min(Lold, a)
• Outgoing rate is the minimum of incoming and the
  fair rate

13
Weighed CSFQ

• If priority is needed for certain specific flows
  this algorithm can be updated to
• include the weighs of the flows and thus the
  dropping probability of the packet would be
• Max(0,1- a wi/r(i))
• The sum of all weigted flow should be a solution
  for the link capacity.
• ?wi min(a,r(i)/wi) C

14
Performance bounds for weighed CSFQ

• During a time interval a flow can receive
  excessive service due to its weight.
• The excessive service can be restricted to this
  flow by implementing a performance bound such
  that it cannot exceed the fair share.

15
Choice of K for algorithm

• e-T/K
• A smaller K increases the system responsiveness
  to rapid rate fluctuation
• A larger K better filters the noise and avoids
  potental system instability The K should be
  larger then delay jitter of the flow and less
  then average duration of the flow

16
Simulations

• Compare the CSFW performance against 4 other
  algorithms
• FIFO and RED do not attempt to achieve fair
  bandwidth allocation
• FRED and DRR Use other approaches to achieve
  fairness

17
A single 10 Mbps Link shared by N flows
Throughput for 32 UDP flows. The arrival rate for
flow is (i1) times larger then its fair share.
18
A single congested link with1 UDP flow and 31 TCP
flows
Throughput when the UDP flow is sending at 10
Mbps and the rest are TCP averaged over a 10
second interval
19
Normalized bandwidth of a TCP flow competing
against N-1 UDP flows. The UDP flows sending at
twice their allocated rates as 210/N
Results DRR performs well when there are less
then 22 flows. CSFQ performs better with TCP
burstiness as it allows packet buffering . CSFQ
performs better then FRED throughout.
20
Multiple congested Link
TCP Sink
Gateway
TCP Source
Gateway
Udp-10
UdpK1
UdpK10
Udp1
21
UDP and TCP flow as a function of no of congested
links and Normalized throughput
a) The flow is UDP
b) The flow is TCP
The UDP and TCP flow send at its fair share
.909Mbps The links are congested since all UDP
send at a rate of 2MBps causing congestion which
has the capacity of 10 Mbps
22
Coexistence of different adaptation schemes

• 3 Receiver layered Multicast flows along with a
  TCP of each having a fair share of 1 Mbps

FIFO
DRR
FRED
RED
CSFQ
FRED in this does not provide fair bandwidth
allocation as RLM and TCP use different end to
end congestion control algorithms
23
On Off Flow with 19 competing TCP flows
Algorithm Delivered Dropped
DRR 601 6157
CSFQ 1680 5078
FRED 1714 5044
RED 5322 1436
FIFO 5452 1306
The On Off perios is derived from exponential
distribution of mean of 100 and 1900 ms. During
the On-Off period the source sends at 10Mbps The
DRR and CSFQ show that the ON-Off source drops
packets since it is limited to its fair share
24
Simulation of web traffic Interarrival times
are exponentially distributed with mean of .05ms
and length of each transfer is from a Pareto
distribution.
Algorithm Mean time Std. dev
DRR 25 99
CSFQ 62 142
FRED 40 174
RED 592 1274
FIFO 840 1695
The mean transfer time and std deviation for 60
short TCP flows in the presence of UDP flow
that sends at link capacity of 10 Mbps Result
CSFQ worse then FRED since it has a larger queue
size.
25
Large Latency model
Algorithm Mean Std. dev
DRR 6080 64
CSFQ 5761 220
FRED 4974 190
RED 628 80
FIFO 378 69
The mean throughputs and std deviations for a
100 second interval for 19TCP and 1 UDp flow at
link capacity of 10 Mbps. The propagation delay
is 100 ms Due to propagation delay we have to
set the buffer size to be 256Kb and K,Ka,Kc to be
400 ms
26
Packet relabeling
Flow1 10 Mbps
Sink
10Mbps Link
10 Mbps Link
Gateway
Flow2 10 Mbps
Gateway
Flow3 10 Mbps
27
Throughput from CSFQ averaged over 10 seconds for
the 3 flows along link 2
Traffic Flow 1 Flow 2 Flow 3
UDP 3.36 3.32 3.28
TCP 3.43 3.13 3.43
28
Identification of Friendly and unfriendly flows
by RED and CSFQ
Simulation1
Simulation2

Algorithm UDP TCP1 TCP2 TCP1 TCP2
RED .906 .280 .278 .565 .891
CSFQ .554 .468 .478 .729 .747
The throughput of one UDP and 2 TCP flows along a
1.5 Mbps link under RED and CSFQ. In first
simulation 1 Mbps UDP flow and 2 TCP flow In the
second simulation we have modified TCP which
opens its congestion window 3 times faster
29
Achieving Fair bandwidth Allocation

• Identifying Flows
• Use a fair detection algorithm to identify
  unfriendly flows
• Isolating Flows
• Router allocates resources separately for
  friendly and unfriendly flows.

30
Summary

• FRED requires per packet flow classification
  while CSFQ does not, so the fairness can be
  achieved in a scalable manner.
• The CSFQ would comprise of high speed backbones
  and lower speed edge routers where classification
  and per flow operation is done.
• The edge routers must be distinguishable from
  the core by some configuration.

31
Acknowledgements

• Professor Bob Kinicky WPI for valuable
  information and advice
• Ion Stoica CMU Author
• Scott Shenker Xerox Author
• Hui Zhang CMU Author
• Nagaraj Shirali For help with Jpeg
  graphs
• Choong-Soo Lee For help with Jpeg graphs