1 Advanced Transport Protocol Design Nguyen Nguyen Multimedia Communications Laboratory March 23, 2005
Jan 20, 2016
1
Advanced Transport Protocol Design
Nguyen Nguyen
Multimedia Communications Laboratory
March 23, 2005
2
Outline
Introduction Overview of TCP/IP System model Queueing model for congestion Loss discrimination Modified AIMD Future work
3
Introduction
Transport protocol End-to-end data transmission Sequencing, flow control, congestion control etc.
Transport protocol measures of performance Throughput (bytes/second) Fairness Latency TCP-friendliness
4
Introduction (2)
Goal: design a reliable transport protocol that achieves high throughput and fairness Key: congestion control
Congestion control design Congestion detection (loss discrimination) Response to congestion
5
Overview of TCP/IP
Layered network architecture
6
Overview of TCP/IP (2)
Internet Protocol (IP) End-to-end data transmission Routing Best-effort
User Datagram Protocol (UDP) Basically raw IP Fast, but unreliable
7
Overview of TCP/IP (3)
Transmission Control Protocol (TCP) Connection-oriented Not as fast as UDP, but reliable
Sliding window transmission policy
8
Overview of TCP/IP (4)
Reliability through retransmission Retransmit lost packets (triple duplicate ACK or
timeout)
Flow control Buffer advertisements from the receiver
Congestion control Congestion indicator = packet loss
9
Overview of TCP/IP (5)
Additive increase,
multiplicative decrease (AIMD)
Additive increase
Triple-duplicate ACK
Timeout
Slow-start
Win
dow
Time (RTT)
10
Overview of TCP/IP (6)
Problem 1. Inaccurate congestion indicator Packet loss in wireless networks is mainly due to
random transmission error (i.e. fading)
Problem 2. Response to congestion TCP is too conservative because it does not have
an up to date notion of the available bandwidth
11
System model
Network model
source1
Internet
source2
destination1
destination2
BS
MH
MH
12
System model (2)
Adjust the rate of the sender subject to the following constraints
Hybrid wired/wireless network topology No help from intermediate routers Unsynchronized clocks Online
13
Queueing model
Single-server queueing system
Customer: packet from primary source plus preceding cross-traffic
Cross-traffic
Primary flow
14
Queueing model (2)
{X2, X3, …} - sequence of interarrival times
{S1, S2, …} - sequence of service times {Q(t) : t ≥ 0} - number of customers in queue
Traffic intensity: ratio of average service time to average interarrival time
][
][
XE
SE
15
Queueing model (3)
D/G/1 queueing system
Case 1: Independent, identically distributed (IID) service times {S1, S2, …} is a sequence of IID r.v.’s
Theorem 1. Let Dn be the departure time of the nth customer. Then {Q(Dn) : n ≥ 1} is a Markov chain.
16
Queueing model (4)
Proof. Let Un be the number of customers arriving during the service time Sn+1 of the (n+1)th customer.
But Un = T-1Sn+1 and service times are independent.
0)(
0)(1)()(
1 DQUDQDQU
DQnn
nnn
n if
if
17
Queueing model (5)
Case 2: Dependent service times {S1, S2, …} is a stationary, ergodic process
{Q(Dn) : n ≥ 1} is not a Markov chain
Theorem 2 [Grimmett]. The waiting time distribution, P(W ≤ w), is non-defective if (a) ρ < 1, or (b) ρ = 1 and Var(S – X) = 0.
18
Queueing model (6)
Long-term properties Average number of customers in the system
Average number of customers in queue, average delay through the system, average waiting time can also be derived
2
][
])[1(2
)(][
1
1
2 SET
SET
SVarTQE
19
Loss discrimination
Improved congestion detection Packet loss Delay
Theorem 2: Long-term stability achieved if average service rate > average arrival rate
The long-term does not exist in our problem
20
Loss discrimination (2)
Sample traffic intensity
Step 1. Calculate the “short-term” average over a time interval
tt ii
ii
i1
1
N
i
s
i
s
N 1
)()( 1
21
Loss discrimination (3)
Condition 1. If > 1 and increasing trend of traffic intensity is observed, congestion
if then
congestion_loss
endif
)(s
)()(
Ms
K
22
Loss discrimination (4)
Condition 2. If a large, sudden spike in traffic intensity is observed, congestion
if then
congestion_loss
endif
)( 1 Mii
23
Loss discrimination (5)
Step 2. Communicate cause of loss to the sender via a feedback message.
Step 3. Retransmit. If cause of loss was congestion, sender adjusts its rate
24
Modified AIMD
Maintain up to date estimate of bandwidth Sample bandwidth
Step 1. Calculate smoothed average
bbb iii)1(
1 1,0,
1
_
ii
i
sizepacketb
25
Modified AIMD (2)
Step 2. Communicate bandwidth estimate to sender via feedback message.
Step 3. Set sending window accordingly
Step 4. Additive-increase.
sizesegment
RTTbwindow
_
26
Future work
Reliability through forward error correction (FEC) instead of retransmission LDPC code Interleaver
Congestion avoidance instead of AIMD