Experiment And Analysis of Dynamic TCP Acknowledgement

1
Experiment And Analysis ofDynamic TCP
Acknowledgement

• Daeseob Lim
• Sam Lai
• Wing-Ho Gordon Wong

2
What is the main problem of dynamic TCP
acknowledgment?

• To ACK, or not to ACK that is the
  question.
• When a packet arrives to the receiver,
  there are two choices
• ACK immediately.
• Wait and ACK later, such that you may have the
  chance to acknowledge multiple packets with just
  one ACK.

3
Whats the difference?

• ACK immediately
• low latency time elapses between the packet
  arrival and the ACK for this packet is send.
• No. of ACK packet increase.
• Wait and ACK later
• High latency
• Small amount of ACK packet generated

4
What consider to be the best solution?

• Low no. of acknowledge
• Low aggregate acknowledgement latency for all
  packets

5
Aggregate latency example
Packet A Arrive
Packet B Arrive
Send single ACK for both A and B
10ms
10ms
10ms
ACK latency for B is 10ms
ACK latency for A is 20ms

• The total ACK latency is 30ms in this case.

6
Naïve Solutions

• Send ACK immediately for each packet?
• Low or even no ACK latency, but this will
  generate too much ACK packets
• Send one ACK for all the end of all
  transmits?
• Only one ACK is needed
• However, high latency
• Dont know which one is the last packet
• What if the link is unreliable?

7
An Online Randomized Algorithm

• Dynamic TCP acknowledgement and other
  stories about e/(e-1)
• By Anna R. Kalin, Claire Kenyon, Dana
  Randall
• Able to achieve a competitive ratio of 1.58
  compare to the optimal solution
• Competitive ratio performance of the
  algorithm / performance of the optimal
  solution

8
Detail about this algorithm

• P(t, t) be the set of packets arrive between
  time t and t
• There exists z such that z is between 0 and 1
  inclusively
• Distributed function to produced z. (randomized
  factor)
• Suppose that ith acknowledgement happens at
  time i and the next one happen at time ti1
• (cont)

9
Detail about the algorithm

• By the algorithm, we should locate Ti1 such
  that ti lt Ti1 lt ti1 and
  P(ti, Ti1)(Ti1 - ti1) z
• If we do that, z unit of latency cost will be
  saved by sending a single additional ACK at Ti1.

10
Before applying the algorithm
11
After the algorithm
12
Why this works?

• The rectangle is guarantee to have area of at
  least 1
• By sending 1 additional ACK, the
  acknowledgement cost increase by 1, but the
  latency cost decreases by at least 1.
• The new sequence is at least as good as the
  original one.
• More detail proof in the paper.
• Dynamic TCP acknowledgement and other stories
  about e/(e-1)

13
Contribution of this research

• Implement a randomized online algorithm about
  delayed ACK into Linux kernel
• Compare real performance of the randomized
  algorithm and the current TCP implementation
• Observe its superiority in terms of cost
• Analyze its inability in terms of throughput

14
ACK for data packet
15
Interval of Delayed-ACK Timer

• Determined by some factors
• Minimum/maximum interval by kernel constants
• Estimated RTT
• Restrictions by RFC 2581
• The maximum is 500ms.
• Acknowledge at least every second segment.
• Acknowledge out-of-order data immediately.
• In most cases, 40ms ( 200ms)

16
Implementation of TCP on Linux
17
Hacking protocol stack
18
Generating random number

• Generate random numbers in advance, store them
  into kernel codes, and select a number
  sequentially

19
Test Environment
NetworkEmulator(ns2)
ModifiedKernelNetworkSniffer(Ethereal)
Router
Client
Server
20
Competitive Ratio Experiment

• The server sends out 100 packets to the client at
  random time spacing at most 70ms apart.
• The competitive ratio is calculated for each cost
  ratio starting from 0.05 to 0.95 stepping by
  0.05, then 0.900 to 0.995 stepping by 0.005.
• Run on simulated networks having bandwidth of
  100Mbps and RTT of 2ms and 100ms for both
  versions of TCP.

21
Overall Competitive Ratioon 2ms Network
22
Overall Competitive Ratioon 100ms Network
23
Blowup Competitive Ratioon 2ms Network
24
Blowup Competitive Ratioon 100ms Network
25
Analysis

• For new TCP, overall the competitive ratio is
  within 1.58 except for borderline cases.
• Small cost ratio
• Expensive latency cost
• Overhead from network sniffer
• Overhead from new TCP
• Large cost ratio
• Expensive acknowledgement cost
• Original TCP acknowledgements
• Possibility of additional acknowledgement

26
Analysis

• For original TCP, the competitive ratio starts
  out extremely high, then converges rapidly with
  the new TCP. Eventually, it starts to increase,
  but at a slower rate.
• Favors delay acknowledgement even when latency
  cost is high.
• Always acknowledge within 200ms or every 2 packet
  full of data even when acknowledgement cost is
  high.

27
Streaming Data Experiment

• The client sends out a request to the server
  asking for data of a certain size to be sent.
• The server replies with the data.
• The client measures the total duration to
  determine throughput.
• Run on simulated networks having bandwidth of
  100Mbps and RTT of 2ms and 100ms for both
  versions of TCP.

28
Streaming Data Result
29
Analysis

• The new TCP can not outperform the original TCP
  in terms of throughput.
• Intuitively, you can imagine, if the incoming
  traffic is regular and data keeps pouring in, to
  optimize throughput, youd want to delay ack as
  long as possible.
• The new TCP can not do delay ack longer than the
  original TCP.
• Random scale down of the threshold for sending an
  additional acknowledgement.
• Our implementation induces little overhead.

30
Conclusion

• Prove the randomized algorithm can achieve the
  competitive ratio of 1.58 in most cases.
• Our implementation achieves better competitive
  ratio comparing to the original TCP in most
  cases.
• Low overhead implementation.
• Can not improve network performance in terms of
  throughput.