Modeling TCP Throughput: A Simple Model and its Empirical Validation

1
Modeling TCP Throughput A Simple Model and its
Empirical Validation

• Ross Rosemark
• Penn State University

2
Goals of Presentation

• Intro to TCP
• Particularly discuss the Reno flavor of TCP
• Show how TCP is innately a self tuning algorithm
• Discuss the goals of this paper
• State assumptions
• Note I do not discus in details any math.
• Discuss Analysis
• State goals of analysis
• State analysis conclusions
• Analyize Paper
• What contributions does it make
• Other these contributions significant?

3
What is TCP?

• According to Computer Networking by James
  Kurose TCP acronym for
• Transmission Control Protocol
• TCP together with UDP form the very core of
  today's Interenet transport layer.
• TCP is a protocol to send messages between two
  machines.

4
So what is the difference between TCP and UDP?

• As discussed, there are two ways to send
  messages..
• UDP
• TCP
• The difference between these protocols is that
• UDP is NOT guaranteed to deliver a packet
• It only does a Best Effort approach
• TCP is guaranteed to deliver a packet

5
How does TCP work?

• When explaining TCP assume that there are two
  nodes (Node A and B) and Node A is sending a
  packet to Node B utilizing TCP.
• High Level example of TCP under ideal
  circumstances
• Node A sets a timer (TimeoutTimer)
• Node A sends a packet (segment) to Node B
• This segment contains a unique sequence number
• Node B sends an Ack to Node A
• Tells Node A that it received the msg.
• Node As TimeoutTimer expires
• Does nothing since received Ack from Node B
• Node B stores the bytes it receives in a buffer
• Application gets bytes from buffer

6
So what are some cases where TCP is not Ideal?

• In TCP a packet can be lost during several
  locations.
• When the packet is transmitted from A to B
• At Node B
• When the packet it transmitted from B to A

7
How can you detect that packets have been lost?

• There are two ways.
• Timeout
• Triple acknowledgement

8
Timeout

• Typically congestion is the reason messages are
  lost due to timeouts

• Node A knows its message is lost if
• Its timer expires and it has not received an Ack
  from Node B
• If a message is lost
• Node A resets its timer
• Node A resends the packet
• Same sequence number
• Process repeats

9
Quick Discussion Based on the discussion. how
is TCP a self tuning algorithm.

• In TCP if node As timer expires before an ack is
  received it sets its timer for double its
  previous timer.
• For instance, if Node A waits 10 seconds for an
  ack for packet 1
• After 10 seconds if node A receives no Ack
• Node A retransmits packet A
• Sets it timer for 20 seconds
• Once a packet an Ack received it sets its timer
  back to the original 10 secs.

10
Triple Acknowledgement

• Another way of determining packet loss
• First things to note
• In TCP packets must be delivered in order
• Also Node A (i.e. sender) can send multiple
  packets at a time
• Technically what is a triple acknowledgement
• When a sender receives multiple Acks for the same
  packet

11
Triple Acknowledgement Example

• Node A sends 5 packets
• Node A never receives an Ack for the second one
• Node B receives packets 1, 3,4, 5
• When node B receives packets 3, 4, 5 send an ack
  that signifies that it never seen packet 2
• Node 2 receives 3 acks for the same packet it
  retransmits lost packet

12
How can you limit packet loss?

• What happens if Node A sends packets quicker then
  Node B processes the packet.
• In this case Node Bs buffer overflows
• To ensure this does not happen TCP implements
• Flow control process
• Speed matching service matching the rate at
  which the sender is sending to the rate at which
  the receiving application is reading

13
How is TCP implement flow control?

• Node B has a buffer
• The available space varies depending on how many
  packets received it received from Node A as well
  as the rate the Node B is processing the packet.
• Every time Node B sends a packet to Node A, Node
  B specifies the remaining size of its buffer
• Node A determines the number of packets it can
  send without overflowing node Bs buffer
• After Node A sends these packets it waits until
  it receives an Ack from Node B.
• Once it receives an Ack Node A can send another
  packet.
• This windows size varies over time.

14
Example of Window Size
15
Does dropped packets affect Congestion Control?

• Oops not only can node B experience congestion
  but routers between Node A and B can experience
  congestion
• How do you adjust the rate A sends packets based
  on these also.
• If Node A experiences packet loss as a result of
  timeout
• it cuts its window size in half.
• It then incrementally increases its window size
  as it receives Acks. (slow start)
• When the window size hits a threshold it grows
  its window exponentially
• If Node A experiences packet loss as a result of
  Triple Ack
• It cuts its window size in half
• It then incrementally increases its window size
  as it receives Acks. . (slow start)
• When the window size hits window size hits ½ the
  value it had before the timeout it exponentially
  grows the window

16
Quick Discussion Again how is TCP a self tuning
algorithm?

• As discussed TCP self tunes to alleviate
  congestion
• I.e. in TCP a nodes window size fluctuates based
  on the ability of Node B to process packets.

17
Reno Flavor TCP

• In this paper they only consider the Reno flavor
  of TCP
• They argue its most widely used on the internet
• Difference between TCP and TCP reno.
• Whenever Node A is in a slow start and receives a
  triple Ack, it cancels the slow start and starts
  growing its window exponentially
• Known as fast recovery

18
Paper

• Paper Title Modeling TCP Throughput A simple
  model and its empirical validation
• Goals of paper
• Model TCP empirically (i.e. through a lot of
  math)
• Predict throughput
• Previous work
• People have previously simulated TCP in
  simulation
• Not sufficient as does not offer TCP-friendly,
  throughput for a non-TCP flow that interacts with
  a TCP connection
• For instance FTPF interacting with TCP

19
Paper

• In this paper the authors state
• Their approach captures
• TCPs fast retransmit mechanism (done in previous
  work)
• TCPs timeout mechanism on throughput (not done
  in previous work)
• Their approach is able to accurately predict
  throughput over a significantly wider range of
  loss rate than before.

20
Assumptions

• They model TCP congestion avoidance behavior in
  terms of rounds
• i.e. round starts with the back-to-back
  transmission of W packets, where W is the current
  size of the TCP congestion widow.
• Once all packets in this window sent no further
  packets are sent until first Ack is received
• The duration of a round is equal to the round
  trip time and is assumed to be independent of the
  window size.
• Also assume that the time it takes to send all
  packets in a window is less then the round trip
  time.
• A receiver sends one Ack for every two packets
  received

21
Assumptions (Cont)

• A packet is lost in a round independently of any
  packets lost in other rounds
• Packet loss is correlated among back to back
  transmissions in a round.
• i.e. one packet in a round dropped all subsequent
  packets in the round dropped.

22
Before we forget lets analyze these assumptions
based on what we know about TCP.

• Are these assumptions acceptable?
• They model TCP congestion avoidance behavior in
  terms of rounds
• Makes sense TCP does the same thing
• The duration of a round is equal to the round
  trip time and is assumed to be independent of the
  window size.
• I would assume the window size does impact the
  round time since the window size dictates how
  many packets are sent..
• Hence this effects the length of the timeout
  timer since the last packet in the windows timer
  will go off before the first packet in the
  widows timer.
• The time it takes to send all packets in a window
  is less then the round trip time.
• No necessarily.. A node could sets its timer to
  short and have to periodically adjust it as it
  further learns the time it takes to send a packet
  to the receiver

23
Assumption Analysis

• A receiver sends one Ack for every two packets
  received
• Sounds good to me I dont know what the problem
  with this would be.
• A packet is lost in a round independently of any
  packets lost in other rounds
• Not valid..
• What happens if node A sends packet 1 to node B
  in round 1 via Router C
• This packet fills Router Cs queue
• Node A realizes this packet is lost and resends
  this packet to B which B receives and sends Ack
  which A receives
• In another round Node A sends packet 2 to node B
  via Router C
• Since packet 1 is still filling up router Cs
  buffer, this packet is dropped.
• As you can see packet1 effects packet2

24
Assumption Analysis

• Packet loss is correlated among back to back
  transmissions in a round.
• Very invalid
• Node A sends in round 1 packet 1 to node B via
  router 1
• Packet lost at router 1
• Node A then sends in round 1 packet 2 to node B
  via router 2
• Packet received by Node B
• Also dont understand how you can get triple acks
  if all subsequent packets are lost?
• Anyone?

25
What do they develop empirical models for?

• Develop stochastic model of TCP congestion
  control in several steps, corresponding to its
  operating regimes
• When loss indication are exclusively from triple
  duplicate acks
• When loss indications are both Timeout and triple
  duplicate
• When congestion window is limited by receivers
  advertised window.
• I dont get into the math at all

26
Measurements and Trace Analysis

• Experimental setup
• Obtained data from 37 TCP connections established
  between 18 hosts scattered across US and Europe
• Operating systems at these hosts varied form
  win95-sun-linux
• Mostly flavors of linux
• TCP varies between approaches
• In linux 2 acks instead of 3 for triple duplicate
  Ack
• All data sets are for unidirectional bulk
  transfers
• i.e. FTP

27
Measurements and Trace analysis

• Example of data utilized in experiments
• Shows overview of traces of machine.
• They would like to point out how Timeout (TO)
  dominated the packets lost.
• This data is used to validate their models

28
Measurements and Trace analysis

• Broke trace data from 1 hour into 100 second
  intervals.
• Each point is a 100 sec interval
• Plotted in lines
• Another papers model
• Considers only triple duplicate acks
• Their model (proposed full)
• TD no timeout only triple duplicate acks
• TO -gt at least one timeout but no exponential
  backoff
• T1 -gt experiences one exponential back off
• i.e. timeout after a timeout
• T2 -gt experience two exponential back off
• T3-gt experience three exponential back off

29
Measurements and Trace analysis

• Their approach better estimates observed model.
• Their approach
• proposed (Approximate)
• Other approach
• TD Only

30
Problems with their model

• Their approach is not always good
• In this graph the other model is a better
  estimator
• They believe because in this data a lot of
  duplicate acks are transpire
• Recall in the other paper they are only trying to
  estimate duplicate acks
• In this paper they are also estimating timeouts..
  Etc

31
Questions?