Alterna(ve Backoff: Achieving Low Latency and High...

17
Alterna(ve Backoff: Achieving Low Latency and High Throughput with ECN and AQM Naeem Khademi , Grenville Armitage, Michael Welzl, Sebas:an Zander, Gorry Fairhurst and David Ros IFIP Networking 2017, Stockholm, Sweden – June 14, 2017 naeemk@ifi.uio.no

Transcript of Alterna(ve Backoff: Achieving Low Latency and High...

Page 1: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

Alterna(veBackoff:AchievingLowLatencyandHighThroughputwithECNandAQM

NaeemKhademi,GrenvilleArmitage,MichaelWelzl,Sebas:anZander,GorryFairhurstandDavidRos

IFIPNetworking2017,Stockholm,Sweden–June14,2017

[email protected]

Page 2: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  Acommonphrasewehaveallheard/said:“Internetistooslowtoday!”

§  FastInternettranslatesintofeelingofcontrolandinterac:vitybytheuser

§  “fast”=>highermediumspeed(bandwidth),shorter:me(delay)andhighertransferrate(transportprotocol)

§  Internetiss:llslowontoday’sbandwidthover-provisionednetworks.Infact,delay(latency)isamajorsourceofprobleminaccesslinks

StuartCheshire(AppleInc.),“It'stheLatency,Stupid”,1996h_p://rescomp.stanford.edu/~cheshire/rants/Latency.html

Internet’sLatencyProblem

2

Source: http://fc04.deviantart.net/fs71/f/2011/330/d/b/slow_internet_by_syas-d4hdtpl.jpg

m3/s t

m3/s

Awater-pipeexample

Page 3: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  WithstandardTCPConges:onControl’sMul:plica:veDecreasefactor(β=0.5):morethroughputmeansmorebuffering(morelatency)o  1BDPforfullu:liza:ono  TCPwillfillupanybuffer!

§  Ac(veQueueManagement(AQM)solu(ons:aimingtokeeptheaveragelatencylow:(FQ_)CoDel,PIE,ARED,…

§  AQMs’defaultthresholdstranslateintoa:nyaveragebuffer(e.g.5ms~20msfor(FQ_)CoDelandPIE)

Internet’sLatencyProblem(#2)

3

7.6 7.8

8 8.2 8.4 8.6 8.8

9 9.2 9.4 9.6 9.8 10

10 20 30 40 50 60 70 80 90 100

Thro

ughp

ut (M

bps)

Buffer Size (pkts)

BDP

2/3 BDP

3/7 BDP

1/4 BDP

1/9 BDP

β=0.5β=0.6β=0.7β=0.8β=0.9

ns-2simula:onvs.modelOneNewRenoflow

@10Mbps,RTT100ms,DropTail

Standard TCP

Page 4: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  Under-u:liza:onbecomesmoreproblema:casRTTfurtherincreases

TheProblemwithUsingAQMs

4

●●

●●●●

●●●●

●●

●●●

●●

●●

●●

Thro

ughp

ut (M

bps) ●

●●

●●●●

●●●●

●●

●●●

●●

●●

●●

20 40 80 160 240RTT (ms)

10

12

14

16

18

20

No_ECN ECN

CoDel

●●

●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●●●●●

●●

●●●●●●●●●

●●

●●●●●●●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

Thro

ughp

ut (M

bps)

●●

●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●●●●●

●●

●●●●●●●●●

●●

●●●●●●●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

20 40 80 160 240RTT (ms)

10

12

14

16

18

20

No_ECN ECN

PIE

One CUBIC flow @ 20 Mbps (real-life tests)

§  Linkunder-u:liza:onwithAQMs’

defaultthresholdswhenRTTsheadaboveintra-countrylevels(60ms~80ms)

Problemwithlong-RTTpaths

link capacity(C.2Tp)

time

Under-utilized portion of the link capacity

buffering (b)

CWmax

β.CWmax

Page 5: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

“CanweusealargerTCPβ?”

5

150

200

250

300

350

0 10 20 30 40 50 60

cwnd

(pa

cket

s)

Time (sec)

Capacity limit

RTTmin 213ms

RTTmin 283ms

Queue length: 1/2 BDPQueue length: 1 BDP

e.g.,CUBICCC(defaultsinceLinux2.6.25)usesβ=0.7

CUBICgeneratesastandingqueue(withDropTail)(i.e.cwndconstantlyabovethecapacitylimit)

Page 6: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  Packetloss:tail-loss(andothersourcesofloss)versusAQM-losso  Notagoodindicatoroftheonsetofconges:on

§  AnECN-CEsignalexplicitlyindicatesanAQM’spresence(mostlikelytobe(FQ_)CoDelorPIE)withalowmarkingthreshold

Alterna(veBackoffwithECN(ABE)

6

§  OnlyasimpleTCPparameterchange=>useaβhigherthan

0.5inresponsetoECN-CE(βECN)andreactwithβ=0.5inresponsetoloss

§  Fewlinesofcodeinthekernel(inthesenderonly)

Alterna(veBackoffwithECN(ABE)

Page 7: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  UsedTEACUP,anexperimentautoma:onsuiteforTCPh_p://caia.swin.edu.au/tools/teacup/

§  TEACUPisdrivingaphysicaltestbedwithadumbbelltopologyusing

FreeBSDandLinuxhosts,andaLinux-basedsotwarerouter(netem)

§  UsedLinux3.17.4implementa:onsofPIEandCoDelwiththeirdefaultparameters

§  iperftogeneratetraffic,tcpdumptocapturethetraffic

§  Per-flowRTTsarecalculatedusingSPPh_p://caia.swin.edu.au/tools/spp/

NoteonExperimentalSetup

7

Page 8: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

Latencyvs.ThroughputTradeoff

8

One NewReno flow, CoDel @ 10 Mbps

0 50 100 150 200 250 300

050

100

RTT=100ms

Byte

s D

epar

ted

βecn

0.60.70.750.8

0.850.90.95

Impr

ovem

ent o

ver β

ecn=

0.5

(%)

0 50 100 150 200 250 300

RTT=200ms

Time(s)0 100 200 300 400 500 600

RTT=400ms

0 2 4 6 8 10

020

4060

8010

0 RTT=100ms

CD

F (%

) βecn

0.50.60.70.750.80.850.90.95

CD

F(%

)

0 2 4 6 8 10

RTT=200ms

Queueing Delay (ms)0 2 4 6 8 10

RTT=400ms

24%improvementforβECN=0.8

Page 9: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

Latencyvs.ThroughputTradeoff(#2)

9

One NewReno flow, PIE @ 10 Mbps

0 50 100 150 200 250 300

050

100

RTT=100ms

Byte

s D

epar

ted

βecn

0.60.70.750.8

0.850.90.95

Impr

ovem

ent o

ver β

ecn=

0.5

(%)

0 50 100 150 200 250 300

RTT=200ms

Time(s)0 100 200 300 400 500 600

RTT=400ms

0 5 10 15 20 25 30 35

020

4060

8010

0 RTT=100ms

CD

F (%

) βecn

0.50.60.70.750.80.850.90.95

CD

F(%

)

0 5 10 15 20 25 30 35

RTT=200ms

Queueing Delay (ms)0 5 10 15 20 25 30 35

RTT=400ms

23%improvementforβECN=0.8

0.7≤βECN≤0.85:agoodtrade-offbetweenimprovedthroughputandincreasedlatency.

Page 10: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

Latencyvs.ThroughputTradeoff(#3)

1010

Thro

ughp

ut (M

bps)

400.5

800.5

1600.5

2400.5

400.6

800.6

1600.6

2400.6

400.7

800.7

1600.7

2400.7

400.8

800.8

1600.8

2400.8

RTT (ms)βecn

5

10

15

20

RTT

(ms)

400.5

800.5

1600.5

2400.5

400.6

800.6

1600.6

2400.6

400.7

800.7

1600.7

2400.7

400.8

800.8

1600.8

2400.8

RTT (ms)βecn

0

50

100

150

200

250 CoDelRTT Throughput

Increased βECN causes insignificant change in median RTT

Source:h_p://frenchyme.blogspot.no

§  Lowlatencyandhighu(liza(on

ABEPerformance

Page 11: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

Latencyvs.ThroughputTradeoff(#4)

11

§  We also tested for One CUBIC flow, with PIE/CoDel @ 10 Mbps

§  Evident throughput gains, although smaller than NewReno’s due to CUBIC’s already aggressive additive increase and default βECN = 0.7

§  Queuing delay distributions are similar to NewReno’s

Page 12: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

TCPConvergenceTime

1212

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5

CD

F

Convergence time / default (βecn 0.5)

βecn 0.6βecn 0.7

βecn 0.75βecn 0.8

βecn 0.85βecn 0.9

Two NewReno flows, CoDel

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5

CD

F

Convergence time / default (βecn 0.5)

βecn 0.6βecn 0.7

βecn 0.75βecn 0.8

βecn 0.85βecn 0.9

Two NewReno flows, PIE

@ {1, 5, 10, 20, 40, 100} Mbps and RTT={10, 20, 40, 60, 80, 160, 240, 320} ms

JFI of the cumulative number of bytes transferred per flow from the time a new flow entered, (after 30 sec during a 90 sec experiment) and noting the time it took until this index reached a threshold (0.95)

Page 13: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

FriendlinesswithStandard(Loss-based)TCP

1313

0.5

1

1.5

2

2.5

3

3.5

1 vs 9 3 vs 7 5 vs 5 7 vs 3 9 vs 1

Nor

mal

ized

thro

ughp

ut ra

tio(E

CN

flow

s / n

oEC

N fl

ows)

# noECN flows vs # ECN flows

βecn=0.5βecn=0.6

βecn=0.7βecn=0.75

βecn=0.8βecn=0.85

βecn=0.9

CoDel

0.5

1

1.5

2

2.5

3

3.5

1 vs 9 3 vs 7 5 vs 5 7 vs 3 9 vs 1

Nor

mal

ized

thro

ughp

ut ra

tio(E

CN

flow

s / n

oEC

N fl

ows)

# noECN flows vs # ECN flows

βecn=0.5βecn=0.6

βecn=0.7βecn=0.75

βecn=0.8βecn=0.85

βecn=0.9

PIE

Normalized throughput ratio between two groups of 10 flows (CUBIC with different βECN against standard TCP) @10 Mbps using RTTpath={20, 80, 160, 320} ms.

For βECN between 0.7 and 0.85 the average throughput ratio is between 1.25 and 2

Page 14: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

ABEandMedium-sizedFlows

14140 2 4 6 8 10 12

010

020

030

040

0

Time (s)

cwnd

(pac

kets

)

RTT, βecn

160ms,0.5240ms,0.5160ms,0.9240ms,0.9

160ms X 20Mbps capacity

240ms X 20Mbps capacity

Theeffectofovershootattheendofslow-start@20Mbps(experiment)

-  Benefitsmedium-sizeflowsi.e.flowstermina:ngrightaterSS-  E.g.reduc:oninPLToflargeweb-pages-  Numberofreduc:onstepsboundtologβECN (0.5)

Lucky flow Unlucky flow

Page 15: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  ABEisbeingconsideredforstandardiza:onasanExperimentaldratbytheInternetEngineeringTaskForce(IETF)

drat-iev-tcpm-alterna:vebackoff-ecn§  ItisdefinedwithinthescopeofECNexperimenta:ondrat

drat-iev-tsvwg-ecn-experimenta:on

§  ItisaWorkingGroupitematTCPMaintenanceandMinorExtensions(tcpm)WorkingGroupattheIETF

§  KernelpatchesforLinuxandFreeBSD,technicalreportandthepaperareavailableath_p://heim.ifi.uio.no/naeemk/research/ABE/

ABE’sStatus

15

Page 16: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

§  ABEisaminorsender-sidemodifica:on(changesβECNonly!)

§  ComplieswiththerestofRFC3168

§  Incrementaldeploymentwithnoflag-day!

Conclusion

16

a)  Significantthroughputgainwithlightly-mul:plexedtraffic

b)  Lowlatency(usingCoDelorPIE)

c)  Nostarva:onofstandardTCP

d)  Reasonableconvergenceandfairnesswithrecommendedβs

ABEPerformance

Page 17: Alterna(ve Backoff: Achieving Low Latency and High ...heim.ifi.uio.no/michawe/research/projects/abe/abe-networking17.pdf · Latency vs. Throughput Tradeoff 8 One NewReno flow, CoDel

Q&A17