Rice Networks Group Aleksandar Kuzmanovic & Edward W. Knightly TCP-LP: A Distributed Algorithm for Low Priority Data Transfer

Rice Networks Grouphttp://www.ece.rice.edu/networks

Aleksandar Kuzmanovic &

Edward W. Knightly

TCP-LP: A Distributed Algorithm for Low Priority

Data Transfer

A. Kuzmanovic and E. W. Knightly

Motivation

Traditional view of service differentiation:– High priority: real-time service– Best-effort: everything else

What’s missing?– Low-priority (receiving only excess bandwidth)– Lower than best-effort!

Non-interactive apps, bulk download Speeds up best-effort service Inference of available bandwidth for resource selection

Routers could achieve via a low (strict) priority queue

Objective: realize low-priority via end-point control– Premise: routers will not help


Applications for Low Priority Service

LP vs. rate-limiting:– P2P file sharing

Often rate limited Isolation vs. sharing

LP vs. fair-share:– Bulk downloads

Improve my other

applications Data-base replication

across the Internet

T im e

R a te lim it

C a pa c ityA va ila bleba ndw idth

G a in

AB


Problem Formulation & Design Objectives

Low-priority service objectives– Utilize the “excess/available” capacity

What no other flows are using

– TCP-transparency (non-intrusiveness)– Inter-LP flow fairness (fair-share of the available

bandwidth)

Design end- host- based transm ission protocol that em ulates the low- priority service


Origins of the Available Bandwidth

Why is excess bandwidth available when TCP is greedy?

1. TCP is imperfect Cross-traffic burstiness Delayed ACKs due to reverse traffic frees up available

bandwidth

2. Short-lived flows Majority of traffic consists of short-lived flows (web

browsing) Bandwidth gaps between short lived-flows


Illustration of TCP Transparency

LP flow utilizes only

excess bandwidth– Does not reduce the

throughput of TCP flows

A B

T C P

LP T im e

F ull C ap ac ity

T C P d e m a n d

T im e

F ull C ap ac ity

L P d e m a n d

T im e

F ull C ap ac ity

B a n d w id th s h a re

L P flo w u tilize sa v a ila b le b a n d w id th

T C P -tra n s p a re n c y


How Is This Different from TCP?

In presence of TCP

cross-traffic:– TCP achieves fairness– LP achieves

TCP-transparency

A B

T C P 1

T C P 2 T im e

F ull C ap ac ityT C P 1 d e m a n d

T im e

F ull C ap ac ityT C P 2 d e m a n d

T im e


In ter-T C P -fairn es s


Fairness Among LP Flows

Inter-LP-fairness is

essential for

simultaneous– file transfers– estimates of available

bandwidth

A B

T C P

LP 1LP 2 T im e

F ull C ap ac ityT C P d e m a n d

T im e

F ull C ap ac ityL P 1 a n d L P 2 d e m a n d s

T im e


In ter-L P -fairn es s


TCP-LP:A Congestion Control Protocol

Key concepts– Early congestion indication

One-way delay thresholds

– Modified congestion avoidance policy Less aggressive than TCP

Implication: Sender-side modification of TCP incrementally deployable and easy to

implement


Early Congestion Indication

For transparency, TCP-LP must know of congestion before TCP

Idealized objective: buffer threshold indication

Endpoint inference: one-way delay threshold

– RFC1323 Source - destination time stamping Synchronized clocks not needed

– Eliminates bias due to reverse traffic

b uffer thres ho ld

)( minmaxmin dddsdi


TCP-LP Congestion Avoidance

Objectives: LP-flow fairness and TCP transparency

LP-flow fairness– AIMD with early congestion indication

Transparency– Early congestion indication– Inference phase goals:

Infer the cross-traffic Improve dynamic properties “MD” not conservative enough


TCP-LP Timeline IllustrationC

onge

stio

n W

indo

w

T im e

- Send 1 pkt/RTT- Ensure available x bandwidth > 0


Con

gest

ion

Win

dow

T im e

TCP-LP Timeline Illustration

- AI phase- CWND/2 upon __early congestion xxindication- Inference phase



onge

stio

n W

indo

w

T im e

-2nd CI => CWND=1- Inference phase



onge

stio

n W

indo

w

T im e


Low-Aggregation Regime

Hypothesis: TCP cannot attain 1.5 Mb/s throughput due to reverse cross-traffic

How much capacity remains and can TCP-LP utilize it?

R 1 R 2

TC P-L P

TC P

C = 1 .5 M b/s

cro s s - t ra f f ic


TCP-LP in Action

TCP alone 745.5 Kb/s TCP vs. 739.5 Kb/s TCP-LP 109.5 Kb/s

TCP-LP is invisible to TCP traffic!

R 1 R 2

TC P-L P

TC P

C = 1 .5 M b/s

cro s s - t ra f f ic


High-Aggregation Regime with Short-Lived Flows

Bulk FTP flow using TCP-LP vs. TCP Explore delay improvement to web traffic Explore throughput penalty to FTP/TCP-LP flow

R 1 R 2... ...

... ...

F ile T ra ns fe r

C lie n t Po o lS e rv e r Po o l

re qu e s t

re s po n s e


TCP Background Bulk Data Transfer

Web response times

are normalizedR 1 R 2... ...

... ...

T C P F ile T ra ns fe r


re qu e s t

re s po n s e


TCP-LP Background Bulk Data Transfer

Web response times improved

3-5 times FTP throughput: TCP: 58.2%

TCP-LP: 55.1%

R 1 R 2... ...

... ...

T C P -L P F T P


re qu e s t

re s po n s e


Conclusions

TCP-LP adds a new service to the Internet– General low priority service (compared to “best-effort”)

TCP-LP is easy to deploy and use– Sender side modification of TCP without changes to routers

TCP-LP is attractive for many applications: ftp, web updates, overlay networks, P2P

Significant benefits for best effort traffic, minimal throughput loss for bulk flows

http://www.ece.rice.edu/networks/TCP-LP

Documents

Rice Networks Group Aleksandar Kuzmanovic & Edward W. Knightly TCP-LP: A Distributed Algorithm for Low Priority Data Transfer