Title: Effect of Algorithms that Improve Fairness of TCP Congestion Avoidance on Performance of Slow links
1Effect of Algorithms that Improve Fairness of TCP
Congestion Avoidance on Performance of Slow links
and Long Thin NetworksVenkatesh Obanaik,
Lillykutty Jacob, A L Ananda,Center for Internet
Research,National University of Singapore.
-
ICCCN 2002 -
Date 14th October 2002
2Contents
- Issues with fairness
- Unfairness of TCP Congestion Avoidance
- Algorithms that improve the fairness of TCP
congestion avoidance - Issue addressed by the paper
- What effect do the fairness algorithms have
on the performance of slow links and LTNs? - Possible solutions
3Fairness
- What is fairness?
- If N TCP sessions share a bottleneck
link, then each session should get 1/N of the
link capacity.
4Unfairness of TCP Congestion Avoidance
- During congestion avoidance phase, cwnd cwnd
1/cwnd, on receipt of every ACK. - TCP sender increases cwnd by utmost 1 segment
after each RTT. - Connections with long RTT open up their windows
relatively slower than connections with short
RTT. - Short RTT connections get more than a fair share
of bottleneck link capacity.
5Algorithms that Improve fairness of TCP
congestion Avoidance
6Constant Rate Policy
- Say, r is the average RTT of the connection.
- During congestion avoidance, the cwnd is
increased at the rate of approximately 1 segment
every r seconds. - Throughput 1/r segments/s in every r seconds
- Rate of increase in throughput is 1/r2
segments/s2 - CR policy suggested an increase of cr2
segments i.e, cwnd cwnd (cr2) / cwnd. - Issues with proper choice of c
7Increase By K Policy
- Describes the standard algorithm as Increase by
1 policy. - Hence IBK implies that, cwnd should be
increased by utmost K segments after every RTT. - Designed for long RTT connections to increase
their throughput without co-operation from other
connections. - To be enabled selectively on long RTT
connections.
8CANIT (Congestion Avoidance Normalized Interval
of Time)
- Addresses the problem in a slightly different
perspective. - In the standard congestion avoidance policy, the
cwnd is increased on the arrival of an ACK
packet. - In short RTT connections ACKs arrive relatively
faster than long RTT connections. - Increase the cwnd of all connections by same
amount in a specified time interval (NIT) - cwnd cwnd (RTT)/(NIT) (1/cwnd)
9 Implicit assumptions in the fairness algorithms
and their effects
- Long RTT of the connection is due to the presence
of a long pipe accompanied by bandwidth - ( Presence of Long thin
Networks not considered ) - Bottleneck is in the core of the network
- ( Presence of Slow access
links not considered ) - Slow links/LTNs only offer a long RTT but not
high bandwidth - TCP sender usually unaware of the network path,
will cause increased probing in a quest for
non-existent bandwidth
10Simulation Study
- Commonly used test configuration in similar
studies conducted previously. - Bandwidth and delay for router R2 and last-hop
router chosen based on a similar study (RFC2416) - MTU size chosen as 296 (according to
recommendations of RFC 3150) - Limited buffer size on last-hop router (RFC 3150
RFC 2757)
11Behavior of IBK,CANIT and CR policies on
connections traversing Slow links/LTNs.
12Variation of congestion window when fairness
algorithms are enabled
- Amount of increase in cwnd caused by each
arriving ACK when different policies are enabled. - CANIT (RTT/ NIT) (1/cwnd)
- (optimum value for NIT
30ms) - CR (cRTT2) / cwnd
- (Value of c 10)
- IBK K/cwnd
- (Value of K 4)
13Variation of cwnd on slow link connection
-
- CANIT Vs
Standard Algorithm
IBK Vs
Standard Algorithm
14Variation of cwnd on LTN connection
- CANIT,CR Vs Standard Algorithm
IBK Vs
Standard Algorithm
15Effect on Performance of Slow links and Long Thin
Networks
- Slow links and LTNs are last-hop links connecting
end-user to the high speed internet - Limited buffer size on last-hop router (ISPs),
sometimes as low as a buffer of 3 packets - Increased last-hop router buffer overflows seen
when the fairness algorithms are enabled on slow
link/LTN connections - Losses very high, especially in the fairness
algorithms that consider RTT as a parameter
16Slow link connection Goodput and Losses
- Goodput Vs QueueSize
Loss Vs QueueSize
17LTN connection Goodput and Losses
- Goodput Vs QueueSize
Loss Vs QueueSize
18Possible solutions
- Increasing last-hop router buffer size
- Improvement in goodput of connections seen
but it also increases the queuing delay and
thereby the RTT -
19Possible Solutions
- Advertising a smaller receive window for slow
link/LTN connections (RFC 3150) -
- Amount of data injected into the
- network reduced, so are the losses.
-
- Only slight improvement seen in
- goodput.
- The sender is limited from
- transmitting new segments during
- fast recovery phase.
20- Inappropriate to apply the fairness algorithms to
slow link/LTN connections.
21Possible Solutions
- Selectively disabling the algorithms on slow
link/LTN connections. - Relatively high goodput and very low losses.
- Since competing connections have alternate
fairness algorithms enabled.
22Conclusions
- Fairness algorithms harm the slow link / LTN
connection. - Inappropriate to apply the fairness algorithms to
slow link/LTN connections. - We recommend selectively disabling the policy on
slow links/ LTNs - The bottleneck is not in the core but is in the
access links.
23Future Work
- Implement mechanisms to disable fairness
algorithms - Test the implementation on the testbed
- Repeat the experiments on the testbed
- Repeat tests on the real network (internet2 link)