Title: A%20Stochastic%20Model%20of%20TCP%20Reno%20Congestion%20Avoidance%20and%20Control
1A Stochastic Model of TCP Reno Congestion
Avoidance and Control
- Jitendra Padhye
- Victor FiroiuiDon Towsley
- Presented by Group 2
2Agenda
- Introduction
- Goals
- Issues Discussed
- Significance of the paper/Relation to Coursework
- Stochastic Model
- Previous Work - Whats different ?
- Issues not Addressed
- Related Work
- Questions/Comments
3Group 2 - Members
- Bijendra Vishal
- Domin Lee
- Guoqiang Shu
- Hannafi Tjan
- Lei Chai
- Jilani Sarmad Syed
- Juan Torres
- Rohit Shenai
- Wissam El Riachy
- Yaoyao Gu
4Tasks/Responsibilities
- Task 1
- Report
- What issues does the paper address ?
- Task 2
- What is the significance of the paper ?
- How is the paper related to the course work?
- Task 3
- Solutions to the Issues mentioned
- Task 4
- How does the paper differ from its previous work?
- Task 5
- What related issues do the paper not address
- Task 6
- Presentation
5Introduction TCP Reno
- Most modern TCPs are Reno
- Reno refined four key mechanisms of TCP
- slow start
- Congestion Avoidance
- Fast Retransmit
- Fast Recovery
6Goals
- Validate the previous model
- Analyze the performance of bulk transfer TCP
flows using more precise, stochastic analysis - Compare the results from the approximate
(previous) model and new and more precise model - Provide both an approximate formula and more
accurate stochastic model for the steady state
throughput of a bulk transfer TCP flow
7Issues Discussed
- Use of a finite state Markov Chain
- For more precise and stochastic analysis
- Probability transition matrix
- Comparisons
- To validate the approximate (previous) model
- Send rate
- Throughputs
8Significance of this paper
- Addresses achievements and weaknesses of previous
work in the area. - Develops a stochastic approach to the problem to
validate a previous and less complicated model. - Extends the knowledge of the topic by developing
a random (real) model for TCP congestion
avoidance.
9Relation to Course Work ????
- Paper uses many of the TCP characteristics
learned in class, (i.e. window size increment
when no losses present). - Connection oriented nature of TCP (concept
learned in class) is clearly seen because a round
finishes when one or more ACKs are received, and
repeated versions of ACKs (at the sender) tell
that a package or more was lost (retransmission
needed).
10Impact of Packet Losses
Packet and ACK transmissions preceding a loss
indication
11Comparison of Throughputs (Cont)
12Analysis of TD period
13Stochastic Model
- To model the behavior of TCP in terms of rounds,
the following random variables are defined - Wi is window size for round i, i 0,8
- Ci allows to model the increment of window by
one every two rounds during no-loss period. - Li is the number of packets lost in the (i-1)st
round, i 0,. 8
14Stochastic Model (Continued)
- Ti denotes whether the connection is in a
timeout state in round i, i0,. 8 - Ri denotes the duration of round i, i 0,..
8RTT denotes the RoundTrip TimeTo denotes the
base timeout value - Ni denotes the number of packets in round i, i
0, 8 - Mi denotes the number of packets transmitted
in round i, i 0, 8
15Stochastic Model (Continued)
- The sequence of random variables
- is finite state Markov chain with probability
transition matrix - Set S is defined as the set consisting of the
states of irreducible sub-chains and over which
the following probability could be defined
16Conditional Expectations
- Let and
denote two successive states in the sequence - Let
- conditional expectations is defined as
17Steady State Send Rate (SR)
- For t gt0, Nt Number of packets transmitted in
the interval 0, tMt Number of packets
received - The long term steady state send rate of a
TCP connection is defined as -
18Steady State Throughput
- The long term steady state throughout of a TCP
connection is defined as
19No Packets are lost
20One or more packets are lost in a round
21One or more packets are lost in a short round
22Exponential BackOff
23Previous Work - Whats different ?
- A simple approximate model was proposed earlier
- SIGCOMM 98
- Predict steady SR as a function of loss rate and
RTT - This paper uses more precise Markov-Chain model
- Result matches the approximation of previous
paper - Verifies the previous results
- Difference from other works (e.g. Ott)
- Model as many as 6 backoff (26 TO)
- Do not model slow start phase
- Do not model fast retransmit
24Comparison of Send Rates
- Approximate model proposed in 9
25Comparison of Send Rates
26Comparison of Throughputs
27Related Issues not Addressed
- Bulk transfer is analyzed
- Short lived flows like HTTP or Real time traffic
may behave differently and may not be captured
the model which is presented in the paper. - Only stochastic analysis is done that has the
following shortcomings - Transition probabilities from one state to
another is assumed to be constant which in real
traffic may vary over time. - Transition probabilities are determined by
present state, not by history - Other characteristic of TCP like fast recovery
and non delayed ACKs are modeled.
28Related Works
- Approaches to analyze TCP connection
- Deterministic analysis of steady state
- Stochastic analysis of the steady state
- This paper
- Fluid queuing model
- Algebraic computation
- Refined models of losses
29Questions/Comments