1. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
A Detailed and Accurate Closed Queueing Network Model of Many Interacting TCP Flows
Michele Garetto
Renato Lo Cigno (presenting)
Michela Meo
Marco Ajmone Marsan
Telecommunication Networks Group Dipartimento di Elettronica - Politecnico di Torino Torino - Italy
Infocom April 26, Anchorage AK, USA

2. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
WHY ?
TCP drives the performance of the Internet
Analytical models help understanding networks and protocols
Good analytical models give insight
Flexible analytical models are required for network planning
Infocom April 26, Anchorage AK, USA

3. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
HOW ?
Decouple the description of the protocol from the descriprion of the network
Find a method that requires only primitive network parameters
Devise a model whose solution complexity is independent from the number of interacting flows
Infocom April 26, Anchorage AK, USA

4. Protocol-network models decoupling
offered
load
TCP
Model
Network
Model
packet loss
probability
RTT
Iterate with a Fixed Point Solution
Infocom April 26, Anchorage AK, USA

5. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
Network Model
Kept as simple as possible, but can be refined at will (if the solution complexity allows!)
FIFO, DropTail queueing
M/M/1/B model, returns
average loss probability
buffer occupancy --> RTT
Infocom April 26, Anchorage AK, USA

6. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
TCP Model
Start from the Finite State Machine (FSM) of the protocol (any version)
Associate an M/M/ queue to every state
Number of customers in the queue is the number of connections in the same state
Customers move from one queue to another following the protocol dynamics
Infocom April 26, Anchorage AK, USA

7. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
TCP Model
The load offered to the network (in packets/s) is computed starting from packets generated in each queue (i.e., state) and queues service times (how long connections stay in each state)
Constant number of concurrent connections (N) yielding a closed queuing network
The solution is simple: solve flow equations balance (independent from N!)
Infocom April 26, Anchorage AK, USA

8. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
An example: SimpTCP
Simplified TCP with
slow start only
maximum window equal to 4 segments
packet loss detection by timeout only
2 exponential backoff before closing
Infocom April 26, Anchorage AK, USA

9. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
SimpTCP
queues representing protocol states
W=1
W=2
W=3
W=4 T0 T1 T2
close
Infocom April 26, Anchorage AK, USA

10. s takes care of ACK spreading
SimpTCP
service times t in queues
W=1
W=2
W=3
W=4 T0 T1 T2
close
t=RTT
t= s RTT
t=RTT
t=RTT
s takes care of ACK spreading
t=t0
t=4t0
t=2t0
t=R.T.
Infocom April 26, Anchorage AK, USA

11. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
SimpTCP
offered load L in packets
W=1
W=2
W=3
W=4 T0 T1 T2
close
L=1
L=2
L=2
L=4
L=1
L=1
L=1
L=0
Infocom April 26, Anchorage AK, USA

12. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
SimpTCP
transition probabilities ps=Tx success
W=1
W=2
W=3
W=4
ps4 ps
ps2
ps2
1-ps
1-ps2
1-ps2
1-ps4 T0 T1 T2
close
1-ps
1-ps
1-ps ps ps ps
Infocom April 26, Anchorage AK, USA

13. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
TCP-Tahoe Model
11 types of queues: the number of queues depends on the maximum window
Type E: model slow start
Type L: model congestion avoidance
Types EF & ET: model periods before a fast retransmit or a timeout respectively
Types F, T0 & T: model f.r. and timeouts (losses)
Types R, EK, TK2 & TK3: model retransmissions and the Karn’s algorithm
Infocom April 26, Anchorage AK, USA

14. TCP-Tahoe Wmax=10 Queuing Network
Infocom April 26, Anchorage AK, USA

15. Network topology for validation
Poli LAN: clients
USA WANs: servers R R
bottleneck #1 10 Mbit/s
bottleneck #2 45 Mbit/s R R
GARR-B & TEN WAN: clients & servers
Infocom April 26, Anchorage AK, USA

16. Packet loss estimate bottleneck #1
0.1
model - B=512 - tic 0.5
B=128 - tic 0.5
Packet loss probability
0.01
simulations - B=512 - tic 0.5
B=128 - tic 0.5
0.001
100
200
300
400
Number of connections
Infocom April 26, Anchorage AK, USA

17. Window size distribution
0.20
Window size distribution
sim - 25 connections
model - 25 connections
0.15
sim connections
model connections
Probability density function
0.10
0.05
0.00 10 20 30 40 50 60
Window size
Infocom April 26, Anchorage AK, USA

18. packet loss with both bottlenecks
model
simulation 20 40 60 80
100
120 N1
200
300
400 N2
0.01
0.1
Packet loss probability
Infocom April 26, Anchorage AK, USA

19. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
Additional features
Results for thousands of concurrent connections
Fixing RTT and Ploss return results for specific connections
Easily adapted to different TCP versions (Reno and New-Reno already implemented)
Introducing classes can manage finite size and short lived connections (done!)
Mix different TCP versions in one model
Infocom April 26, Anchorage AK, USA

20. Infocom 2001 - April 26, 2001 - Anchorage AK, USA
Conclusions
New modeling technique based on queuing networks
Powerful: starts from FSM description of protocols and physical parameters of the network
Simple solution
Can be used for planning and dimensioning
Infocom April 26, Anchorage AK, USA