1. Columbia University in the city of New York
Providing Throughput Differentiation for TCP Flows Using Adaptive Two-Color Marking and Two-Level AQM
Presented by
Vishal Misra
Columbia University in the city of New York
Joint work with
Y. Chait, C.V. Hollot, Don Towsley, H. Zhang (UMass-Amherst) and John Lui (Chinese University of Hong Kong)
Infocom 2002

2. Overview Background and motivation Fluid-flow model Simulations
Two level PI
Adaptive Rate Marker
Simulations
Conclusions
Infocom 2002

3. Diffserv Architecture: Background
A,B
marking
End-host:
- negotiates a profile with edge-router
scheduling
Edge router:
- per-flow traffic management
- marks packets as in-profile and out-profile
Core router:
- per class traffic management
- buffering and scheduling
based on marking at edge
Infocom 2002

4. Leaky-bucket Marking at Edge
profile: pre-negotiated rate A, bucket size B
packet marking at edge based on per-flow profile
Rate A B
User packets
Infocom 2002

5. Assured Forwarding at Core
active queue management
computes average queue length, x
p1: drop prob. of green packet
p2: drop prob. of red packet 1
Avg. Queue length
Drop prob
Infocom 2002

6. TCP over AF Service TCP Other flows
Profile:A,B
marker
bottleneck core
TCP
Other flows
Question: is it possible to provide a TCP flow a fixed (minimum) rate through proper choice of parameters (A,B)
Studied in
“Achievable Service Differentiation with Token Bucket marking for TCP” [Sahu et al. Sigmetrics 2000]
Infocom 2002

7. Ideal Differentiation Not Possible
consider two identical TCP flows (f1, f2)
conventional service
same achieved rate for both flows
assured forwarding
ideally want to have achieved rate, r, proportional to assured rate A, i.e,
r1/r2 = A1/A2
not possible with token parameter setting
Profile-based marking favors flows with lower token-bucket rate A
Infocom 2002

8. Proposed Solution Make bucket rate adaptive
A -> A(t)
Design controller to set A(t) at the edge
Implement two-level PI at core
Infocom 2002

9. TCP - Fluid Flow Model [MGT Sigcomm 2000]
Window size
Queue length
Round trip
add
increase
mult
decrease
loss
arrival rate
outgoing
traffic
incoming
traffic
propagation
delay
queuing
delay
Infocom 2002

10. Fluid Flow Model: Adding DiffServ
Token bucket per aggregate:
Fraction of fluid
marked green
Two-color marking
Loss probability
Infocom 2002

11. 2-level PI Controller at Core
Use single controller
Define two desired queue lengths and
PI regulates buffer queue to
For over-provisioned, queue converges to , for under-provisioned queue converges to
Infocom 2002

12. Adaptive Rate Marker at Edge
ARM: Another PI controller
Input signal: (estimate of)
achieved throughput
Output signal: Bucket rate A
Infocom 2002

13. Linearized Fluid Model
Infocom 2002

14. ns-2Simulation topology
Infocom 2002

15. ns Simulations
Smooth curve : fluid model
Infocom 2002

16. ns Simulations
Added transient FTP flows
Added HTTP flows
Infocom 2002

17. ns Simulations Added transient FTP flows Added HTTP flows
Increase capacity by 20%
Infocom 2002

18. ns Simulations
Added another edge w/o SLA
Infocom 2002

19. ns Simulations Added another edge w/o SLA Increase capacity by 20%
Infocom 2002

20. ns Simulations
Decrease capacity by 20%
Infocom 2002

21. Conclusions and Future Work
A modification to the DiffServ architecture proposed
Control theoretic design and stability analysis of system performed
SLA-consistent throughput differentiation simulations
Investigate proportional (or fair) allocation of excess bandwidth
Improve model to account for small bucket size
Infocom 2002