On the Use of Anycast in DNS Sa nde e p Sa ra t Jo hns Ho pkins - - PowerPoint PPT Presentation

On the Use of Anycast in DNS

Sa nde e p Sa ra t Jo hns Ho pkins Unive rsity Va sile io s Pa ppa s UCL A Andre a s T e rzis Jo hns Ho pkins Unive rsity http:/ / hinrg .c s.jhu.e du/

2

Background

• Wha t is Anyc a st?

– Clie nt tra nspa re nt me c ha nism to ro ute pa c ke t to o ne o f multiple se rve rs in a nyc a st g ro up – I mple me nte d via a nno unc e me nts o f the sa me a ddre ss pre fix fro m multiple o rig ins (I GP+E GP) – De plo ye d in to p-le ve l DNS na me se rve rs

• Re duc tio n in q ue ry la te nc y
• Sc a la b ility
• Ava ila b ility
• Re sista nc e to DDo S a tta c ks

3

Goal

• Me a sure the impa c t o f a nyc a st o n

DNS

– Re spo nse time s – Ava ila b ility in te rms o f numb e r a nd dura tio n o f o uta g e s – Co nsta nc y o f se rve r se le c tio n – E ffe c tive ne ss o f lo c a liza tio n

4

What we tested

• Ba se Ca se : Unic a st se rve r

– T e st-c a se : B-Ro o t (lo c a l lo a d b a la nc ing )

• Anyc a st Co nfig ura tio ns

– Hie ra rc hic a l

• T

e st-Ca se s: F

• Ro o t (26 se rve rs), K
• Ro o t (7

se rve rs)

– E xplo re the e ffe c t o f numb e r a nd lo c a tio ns o f se rve rs

– F la t

• T

e st-Ca se : Ultra DNS (8 se rve rs *)

5

Measurement Methodology

• Me a sure me nts using

Pla ne tL a b

• Spe c ia l DNS q ue rie s to

the a nyc a st a ddre ss fro m e a c h PL site e ve ry [25-35] se c o nds

• Pe rio d o f study: 3 we e ks fro m Se pt 19,

2004 to Oc t 8, 2004

• De finitio ns

– Outa g e : Pe rio d o f time whe n q ue rie s a re una nswe re d (multiple o f me a s. pe rio d) – F lip: Clie nt switc he s fro m o ne se rve r to a no the r

6

Response Times

• Anyc a st se rve rs ha ve

lo we r re spo nse time s

• Ultra DNS T

L D1 ha s the lo we st q ue ry la te nc y

• Amo ng the re st, F
• Ro o t is the b e st

– Re a so n: hig h g e o g ra phic dive rsity

• Re spo nse time s ha ve

hig h de via tio ns

– Due to insta b ility a s we will se e la te r

* Hypo the tic a l c a se s pure ly fo r c o mpa riso n

Server Mean Median (ms) (ms) (ms) 45 35 13 min{TLD1.TLD2}* 69 51 173 TLD1 96 54 207 F-Root 75 70 85 TLD2 104 85 237 B-Root 115 95 121 K-Root 140 121 104

• Std. dev
• Hypo. Unicast*

Flat vs. Hierarchical Effect of server location

7

Availability

• Pe rc e nta g e o f

una nswe re d q ue rie s < 0.9%

• T

L D1,T L D2 ha ve the la rg e st numb e r o f

• uta g e s
• F
• Ro o t ha s the le a st

– Re a so ns (spe c ula tio n)

• Ultra DNS is sing le -ho me d
• L
• ng e r I

nte rne t pa ths

• Ave ra g e inte r-o uta g e

time fo r sa me c lie nt is in the o rde r o f da ys

8

Outage Duration

All sc he me s sho w ro ug hly sa me b e ha vio r

9

Constancy

• Co nsta nc y me a sure d

b y fre q ue nc y o f flips b e twe e n se rve rs

• T

L D1, T L D2 ha ve mo st flips

• F
• Ro o t, K
• Ro o t ha ve

hig he r pe rc e nta g e o f flips a fte r a n o uta g e

• Ma jo rity o f flips fo r F
• Ro o t a nd K
• Ro o t a re

b e twe e n the g lo b a l no de s

5 o rde rs o f ma g nitude

10

Effectiveness of Localization

• Que stio n: Do e s

a nyc a st le a d c lie nts to the c lo se st se rve r?

• Dire c t c o mpa riso n

fla we d due to diffe re nt ro uting pa ths fo r unic a st a nd a nyc a st a ddre sse s

• So lutio n:

– Co mpa re pa th use d b y a nyc a st to pa ths to a ll la st ho p ro ute rs 60% to 80% o f c lie nts g o to the c lo se st a nyc a st se rve r

11

Comparison of Strategies

• Hie ra rc hic a l sc he me s ha ve hig he r

sta b ility a nd a va ila b ility

• F

la t sc he me s a re mo re e ffe c tive in dire c ting q ue rie s to the “c lo se st” a nyc a st insta nc e

• Po ssib le ide a :

– T une pa ra me te r to a da ptive ly c ha ng e pro pe rtie s a nyc a st sc he me – Ra dius o f a nno unc e me nt a t e a c h a nyc a st no de

12

Summary

• Anyc a st impro ve s a va ila b ility
• Othe r pro pe rtie s de pe nd o n the sc he me

use d

• T

ra de -o ff b e twe e n a va ila b ility, sta b ility a nd e ffe c tive ne ss o f lo c a liza tio n

• Ca ve a ts:

– Re sults a pply to Pla ne t L a b e nviro nme nt – Suppo rt a rg ume nts using BGP da ta – Ske w due to lo a d o n the a nyc a st se rve r

• F
• r mo re :

– http:/ / www.c s.jhu.e du/ ~sa ra t/ Anyc a st-T R.pdf