Dynamic-CBT and ChIPS - Router Support for Improved

Multimedia Performance on the Internet

Jae Chung and Mark Claypool

Department of Computer Science

WPI

In Proceedings of ACM Multimedia

November 2000

The Internet and Multimedia

• Internet routers are best effort– No timing constraints

– Packet loss, which indicates congestion

• TCP– Completely reliable delivery through retransmission

– Respond to loss as congestion

• But … TCP often unsuitable for interactive multimedia– Full reliability not needed

– Window-based rate fluctuations

Multimedia Using TCP

Multimedia Using UDP

Multimedia on the Internet

• Multimedia often uses UDP– Avoid delay and jitter from retransmission

– Rate-based

– Unresponsive!

• Router queue management goals– Congestion Control

– Fairness

– Reduce Jitter

Current Router Queue Management

Drop Tail (FIFO)Drop Tail (FIFO)

Resource ReservationResource Reservation

...... CBQCBQ Floyd, Jacobson

93 Floyd, Jacobson

95 Lin, Morris

97

Bursty loss

Unfairness (tcp, other)

No QoS support

REDRED FREDFRED

Active Queue MgmtActive Queue Mgmt

......

Class-Based Threshold (CBT) AQM Support for Multimedia - Jeffay, 99

Drop Tail (FIFO)Drop Tail (FIFO)

Resource ReservationResource Reservation

...... CBQCBQ REDRED FREDFRED

Class-Based Threshold (CBT)

Class-Based Threshold (CBT)

Active Queue MgmtActive Queue Mgmt

......

Outline

• Introduction • CBT and D-CBT

– Design

– Evaluation

• ChIPS– Design

– Evaluation

• Conclusion

CBT Concepts

• CBQ + RED: Class-based isolation on RED

• Use Class Thresholds and Avg. # of enqueued packets on a single FIFO Queue

• Three classes: TCP, MM UDP (flow controlled) and Unresponsive UDP

TCP

MMUDP

UDP

RED

CBT Design

CBT - Pros and Cons

• Pros: RED + Class-Based Isolation– Early Congestion Notification

– Protect TCP, and protect (distinguish) MM UDP

– Different flows coexist with predefined fairness.

– Dividing bandwidth assigned is up to the class.

• Cons: CBQ function w/o admission control– Might not work well for certain traffic mixes

– Arguable that it’s not fair (as in the case of CBQ)

Dynamic-CBT

Drop Tail (FIFO)Drop Tail (FIFO)

Resource ReservationResource Reservation

...... CBQCBQ REDRED FREDFRED

Class-Based Threshold (CBT)

Class-Based Threshold (CBT) Dynamic-CBTDynamic-CBT

Active Queue MgmtActive Queue Mgmt

......

ChIPSChIPS

Dynamic-CBT and ChIPS

Drop Tail (FIFO)Drop Tail (FIFO)

Resource ReservationResource Reservation

...... CBQCBQ REDRED FREDFRED

Class-Based Threshold (CBT)

Class-Based Threshold (CBT) Dynamic-CBTDynamic-CBT

Active Queue MgmtActive Queue Mgmt

......

D-CBT Design

Responsive MM

EnableThrsh

Flow Counting in D-CBT

• For every incoming packet, insert or update <dest-addr, flow-id, local-time> info and update count– Sorted Linked List - O(n)– Hash Table - O(1)

• Every ms, delete old info and update count– Sorted Linked List - O(n)– Hash Table - O(n)

• (How are flows counted in FRED?)

Evaluation in NS

• Developed responsive multimedia application (for tagged UDP class)– AIMD Media Scaling (5 discrete rates)

– “MPEG-1 like” transmission rates

– [CC00a], MM-Flow

• Implemented and validated CBT

• Implemented D-CBT and measured congestion time fairness– RED vs. CBT vs. D-CBT

Validation of CBT on NS

0 20 60 110 160 180

6 ProShare - Unresponsive MM (210Kbps each)

240 FTP-TCP

1 UDP blast (10Mbps, 1KB)

• RED Settings:qsize = 60 pktsmax-th = 30 pkts min-th = 15 pktsqweight = 0.002max-pro = 0.1

• CBT Settings:mm-th = 10 pkts

udp-th = 2 pkts

(Second)

X axis: Seconds, Y axis: Kbyte/Sec

PJS99 Experimental Results

NS Simulated Results

Aggregate TCP Throughput under RED

(Our setup is ok, so now can check our CBT test)

Aggregate TCP Throughput under CBT

X axis: Seconds, Y axis: Kbyte/Sec

PJS99 Experimental Results

NS Simulated Results

Simulation (RED, CBT, D-CBT)

0 10 20 30

25 FTP-TCP

30 FTP-TCP

2 CBR-UDP(5Mbps each)

10 MM-UDP

s1

n1

s2

s66

s67

n2

r2

r2

r66

r67

25Mbps, 20ms

25Mbps, 5ms25Mbps, 5ms

n1-n2: q_size = 60

RED: max_th = 15 min_th = 5 qweight = 0.002 max_prb = 0.1

CBT: mmu_th = 2.9 udp_th = 0.6Second

Jain’s Fairness Index ( f ) - Jain, 91

• Examples:– 1 flow– 2 flows, 5 Kbps each– 2 flows, 9 Kbps and 1Kbps

f (x0, x1, x2,…, xn) =( xi )2i=0

n

i=0n xi

2n

0 f 1 (Greatest Fairness)

RED: Class Average Per-Flow Throughput

537293 161

851 720 601

45034681

0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

0 ~ 10 Sec 10 ~ 20 Sec 20 ~ 30 Sec

Kbp

s

TCP

TAG_UDP

UTAG_UDP

Fairness: RED

Fairness: CBT

CBT: Class Average Per-Flow Throughput

587 613

305

760 739657624

318

0

100

200

300

400

500

600

700

800

900

1000

0 ~ 10 Sec 10 ~ 20 Sec 20 ~ 30 Sec

Kbp

s

TCP

TAG_UDP

UTAG_UDP

Fairness: D-CBT

D-CBT: Class Average Per-Flow Throughput

622659

356

653 660

378

637

313

0

100

200

300

400

500

600

700

800

900

1000

0 ~ 10 Sec 10 ~ 20 Sec 20 ~ 30 Sec

Kbp

s

TCP

TAG_UDP

UTAG_UDP

Outline

• Introduction • CBT and D-CBT

– Design – Evaluation

• ChIPS – Design

– Evaluation

• Conclusion

Cut-In Packet Scheduling (ChIPS) Design

ChIPS Evaluation - Jitter

Frame Sequence Number

MM Frame Delay

Sec

ond

D-CBT-FIFO MM1 DelayD-CBT-ChIPS MM1 Delay

ChIPS Evaluation - Fairness

  TCP Packets Delivered

TCP Packet Drop Rate

TCP Throughput

D-CBT 66,648 pkts 4.46 % 17,773 Kbps

D-CBT w/ ChIPS

66,386 pkts 4.44 % 17,703 Kbps

D-CBT: Class Average Per-Flow Throughput

622659

356

653 660

378

637

313

0

100

200

300

400

500

600

700

800

900

1000

0 ~ 10 Sec 10 ~ 20 Sec 20 ~ 30 Sec

Kbp

s

TCP

TAG_UDP

UTAG_UDP

D-CBT with ChIPS: Class Average Per-Flow Throughput

608666

355

673 664

384

563

297

0

100

200

300

400

500

600

700

800

900

1000

0 ~ 10 Sec 10 ~ 20 Sec 20 ~ 30 Sec

Kbp

s

TCP

TAG_UDP

UTAG_UDP

Conclusion

Fair toTCP

Fair toMixed(Fixed)

Fair toMixed

(Variable)

ImproveMM Jitter

D-CBTChIPS X X X X

RED X

CBT X X

D-CBTFIFO X X X

Future Work

• Active Flow Counting (Overhead)

– For every incoming packet, update flow infoHash Table - O(1)

– Every ms, delete old flowsHash Table - O(n)

• Measure Overhead– Processing Time and Memory Usage

Future Work

• How many different classes are needed?– Example

1 class is RED

1 class per flow is FRED

– Overhead per class

• Effects of D-CBT and ChIPS on Perceptual Quality

Evaluation of Science?

• Category of Paper

• Science Evaluation (1-10)?

• Space devoted to Experiments?