Raj JainThe Ohio State University

Raj Jain, Shiv Kalyanraman, Ram ViswanathanDepartment of Computer and Information Sci.

The Ohio State UniversityColumbus, OH 43210

Jain@ACM.Org

Raj JainThe Ohio State University

Features

Scheme

Simulation results

Overview

Raj JainThe Ohio State University

Design FeaturesCongestion Avoidance

High throughputLow delayOperation independent of buffers

LinkUtilization

Time

QueueLength

Time

5

Raj JainThe Ohio State University

Design FeaturesWorks for bursty trafficFairness: Analytically provenMeasured overload/load at switch/sourceBounded oscillations ⇒ Good for VBRParameters: Few, insensitive, easyBipolar feedbackSeveral options: BECN

Precise fairness computation

Raj JainThe Ohio State University

Transmission Rate

The OSU Scheme

The sources periodically send a control cellThe switches measure load over a periodThe switches specify reduction factor in cellThe destination returns the cell to the sourceThe source adjusts the transmission rate

Offered average Rate Adjustment Factor

Raj JainThe Ohio State University

Control Cell FormatTransmitted Cell Rate (TCR)=1/Inter-cell timeOffered average cell rate (OCR)

Load adjustment factor: TCR New = TCR/LAFAveraging intervalFor BECN option: FECN/BECN bitTimestamp (when sent by the source)

1 ms

10 µs

TCR OCR

Raj JainThe Ohio State University

Source AlgorithmArbitrarily select: Initial TCR

Averaging interval TSend sells at inter-cell time = 1/TCRSend control cells every TOn receiving a control cell:

T ← Averaging interval from cellExecute LAF algorithm

Raj JainThe Ohio State University

Load Adjustment Factor in Cell<1Yes No

Current TCR = New TCR

YesNo

Ignore

Yes No

Ignore

New TCR = TCR in Control Cell

Load Adjustment Factor in Cell

New TCR <Current TCR

New TCR >Current TCR

Source LAF Algorithm

Raj JainThe Ohio State University

Switch LAF AlgorithmParameters:

Averaging interval TTarget utilization band (TUB): U(1± ∆)

Compute Target # of cells per T at UCount cells received over TLoad = Received/TargetTcell ← Max{Tcell, T}LAFcell ← Max{LAFcell, Load}

LoadLevel

Time

1

Raj JainThe Ohio State University

The TUB Algorithm

In-TUB?

OCR<Fair share?

LAFcell ← Max{LAFcell, Load/(1-∆)}

LAFcell ← Max{LAFcell, Load/(1+∆)}

LAFcell ← Max{LAFcell, Load}

Fair share = Target rate/# of active VCsAnalytically proven:Fair and Closed (remain in TUB)

Raj JainThe Ohio State University

SecondSource’s

Ratey

First Source’s Rate x

23

4

56

(x, y)

(x’, y’)

1

Proof

Fairness Line

x + y = U(1+∆)x + y = Ux + y = U(1-∆)

Efficiency Line

Raj JainThe Ohio State University

Raj JainThe Ohio State University

Single-Source Configuration

All links 150 Mbps, 1 kmMax-min optimal: 150 MbpsTUB = 0.90 (1 ± 0.1)Averaging interval = 150 µs

Sw 1 D1S1 Sw 1

Raj JainThe Ohio State University

Two-Source Configuration

All links 150 Mbps, 1 kmMax-min optimal: 75 MbpsTUB=0.90(1 ± 0.1)Averaging interval = 150 µs

Sw 1S1

Sw 2S2

D1

D2

Raj JainThe Ohio State University

Three-Source Configuration

All links 150 Mbps, 1 kmMax-min optimal: 50 MbpsTUB=0.90(1 ± 0.1)Averaging interval = 150 µs

S1

Sw 1 Sw 2S2

S3

D1

D2

D3

Raj JainThe Ohio State University

Transient Configuration

All links 150 Mbps, 1 kmTUB=0.90(1 ± 0.1)Averaging interval = 150 µs

Sw 1S1

Sw 2S2

D1

D2

OptimalRate

Time

S1

S2

Raj JainThe Ohio State University

Parking Lot Configuration

All links 150 Mbps, 1 kmMax-min optimal: 50, 50, 50 MbpsTUB=0.90(1 ± 0.1)Averaging interval 150 µs

S1

S2Sw 1 Sw 2 Sw 3

S3

D1

D2

D3

Raj JainThe Ohio State University

Upstream Configuration

All links 150 Mbps, 1 kmMax-min optimal: 50, 50, 50, 100 MbpsTUB=0.90(1 ± 0.1)Averaging interval = 150 µs

Sw 1 Sw 2 Sw 3

S4

D3

D4

S2

S1 D1

D2S3

Raj JainThe Ohio State University

VC Cell Rates

Raj JainThe Ohio State University

Link Utilization

Raj JainThe Ohio State University

Queue Length

Raj JainThe Ohio State University

Queue Size with Startup at PCR/10

Raj JainThe Ohio State University

VC Cell Rates in WAN Configuration

Raj JainThe Ohio State University

Queue Size with WAN Configuration

Raj JainThe Ohio State University

VC rates with Train Traffic

Raj JainThe Ohio State University

BECN Option

TUB Algorithm

AggressiveFairnessOption

PreciseFair share

ComputationOption

Switch LAF Algorithm

Load Measurement at Switches

Load Measurement at Sources

Family Portrait

Source LAF Algorithm

1 2 3 4 65

Raj JainThe Ohio State University

Summary

Design Principles: Input rate overload, Feedback-control relationship, TCR/OCR specification, TUB fairness

Features: High throughput, Low delay, Avoidance,Bipolar

Basic Scheme: Source/switch LAF algorithm

Options: Thoroughly tested all variations