One More Bit Is EnoughOne More Bit Is Enough

Yong Xia, RPI

Lakshminarayanan Subramanian, UCB

Ion Stoica, UCB

Shivkumar Kalyanaraman, RPI

SIGCOMM’05, August 22-26, 2005, Philadelphia, Pennsylvania, USA

VCP

Goal

2

Achieve fair bandwidth allocation and high utilization, and minimize packet loss in high bandwidth-delay product (BDP) network TCP ? TCP + AQM / ECN? XCP ?

VCP 3

Why TCP does not scale?

TCP uses duplicate ACK or timeout: no degree of congestion and instability

time

congestionwindow Multiplicative Decrease (MD)

Additive Increase (AI)

congestion

slow start

congestion avoidance

AI with a fixed step-size can be very slow for large bandwidth

slow!

VCP 4

TCP does not perform well in high b/w

bottleneck utilization

round-trip delay (ms)

TCP

50%

100%

bandwidth (Mbps)

bottleneck utilization

TCP

70%

100%

VCP 5

XCP scales

x

sender receiverrouter

C

spare bandwidth

DATA

ACKrate

But, XCP needs multiple bits (128 bits in its current IETF draft) to carry the congestion-related information from/to network

VCP

Goal

6

Design a TCP-like scheme that:

requires a small amount of congestion information (e.g., 2 bits)

scales across a wide range of network scenarios

General idea

degree of explicit information

TCP: no explicit feedback, relay on duplicate ACKand timeout

ECN: one bit, underload and overload

VCP: two bits, low-load, high-load and overload

XCP: multiple bits, report spare bandwidth

VCP 8

Variable-structure congestion Control Protocol (VCP)

sender receiver

x

router

Routers signal the level of congestion End-hosts adapt the control algorithm accordingly

traffic rate link capacity

(11)

(10)

(01)

code loadfactor

region

low-load

high-load

overload

control

Multiplicative Decrease (MD)

Additive Increase (AI)

Multiplicative Increase (MI)

range of interest

ACK

2-bit ECN

0

1

scale-free

2-bit ECN

VCP 9

VCP vs. ECN

sender receiver

x

router

code loadfactor

region

overload

control

Multiplicative Decrease (MD)

0

1

ECN doesn’t differentiate between low-load and high-load regions

TCP AQM / ECN

(11)

(10)

(01)low-load

high-loadAdditive Increase (AI)

Multiplicative Increase (MI)

ACK

2-bit ECN

VCP

(1)

(0)underloadAdditive Increase (AI)

ACK

1-bit ECN

VCP 10

VCP vs. ECN

TCP+RED/ECN

cwnd

utilization

time (sec)

VCP

utilization

cwnd

time (sec)

VCP 11

VCP key ideas and properties

Achieve high efficiency, low loss, and small queue Fairness model is similar to TCP:

Long flows get lower bandwidth than in XCP (proportional vs. max-min fairness)

Fairness convergence much slower than XCP

Use network link load factor as the congestion signal

Decouple efficiency and fairness controls in different load regions

overload

high-load

low-load

router

fairness control

efficiency control MI

AIMD

end-host

VCP 12

Major design issues

At the router How to measure and encode the load factor?

At the end-host When to switch from MI to AI? What MI / AI / MD parameters to use? How to handle heterogeneous RTTs?

control

Multiplicative Decrease (MD)Additive Increase (AI)

Multiplicative Increase (MI)

(11)(10)

(01)

code load region

low-load

high-loadoverload

VCP 13

Design issue #1: measuring and encoding load factor

Calculate the link load factor

demandload_

factor capacity =

link_bandwidth * t

arrival_traffic + queue_size =

t = 200ms most rtt

The load factor is quantized and encoded into the two ECN bits

VCP 14

Design issue #2: setting MI / AI / MD parameters (, , )

overload

high-load

low-load MI

AIMD

VCP

Design issue #2: setting MI / AI / MD parameters (, , )

load factor

100%

80%safety margin

0%

0.875

15

TCP: = 1.0 VCP: =

0.06STCP: =

0.01

MI

AI

MD

k (1 *) / *

where k = 0.25 (for stability)

80%

20%

1.0 = =

=

* =

VCP

Design issue #3: Handling RTT heterogeneity for MI/AI

16

VCP

VCP scales across b/w, rtt, num flows

Evaluation using extensive ns2 simulations

17

150Mbps, 80ms, 50 forward flows and 50 reverse flows

Impact of bottleneck capacity

vary link capacity from 100Kbps to 5Gbps

high utilization, gap comparing toXCP at most 7%

at extremely low capacitiesalpha = 1 is too large for such low capacity

Impact of feedback delay

fix bottleneck capacity at 150Mbps, vary round-trip propagation delay from 1ms to 1500ms

utilization > 90%

average queue < 5%maximal queue < 15%

no packet losslower utilization due to comparable smallmeasurement interval, i.e. tp = 200ms << RTT

Impact of number of long-lived flows

increase the # forward FTP flows

small queue length, even outperform XCP

high utilization

zero packet drop

Impact of short-lived traffic

arrive according to Poisson process,avg. arrival rate varying from 1/s to 1k/s,transfer size obeys Pareto distribution with avg. 30 packets

high utilization

small queue length

zero packet drop

Multiple bottlenecks

typical parking-lot topology

as good as single-bottleneck scenarios

zero packet drop

< 0.2% buffer size average queue length

Fairness

(a) same RTT(b) small RTT difference(c) huge RTT difference

even distribution among all flows in case (a) and (b)

fairness discrepancy due to high value of MI and AI,whose value is bound in implementation to prevent burst

Convergence behavior

introduce 5 flows one after another

high utilization during the whole period

use router buffer size to scale queue length axis;low queue length during the whole period

Sudden demand change

150 flows join at 80s and leave at 160s

high utilization

remain much lower than full size

VCP

Conclusions

26

With a few minor changes over TCP + AQM / ECN, VCP is able to approximate the performance of XCP

High efficiency Low persistent bottleneck queue Negligible congestion-caused packet loss Reasonable (i.e., TCP-like) fairness

VCP

VCP comparisons

27

Compared to TCP+AQM/ECN Same architecture (end-hosts control, routers signal) Router congestion detection: queue-based load-based Router congestion signaling: 1-bit 2-bit ECN End-host adapts (MI/AI/MD) according to the ECN feedback End-host scales its MI/AI parameters with its RTT

Compared to XCP Decouple efficiency/fairness control across load regions Functionality primarily placed at end-hosts, not in routers

VCP 28

THE END

VCP Parameter Setting

Modified from

www.ecse.rpi.edu/Homepages/ shivkuma/research/papers/vcp05.ppt

Remark