LB

Server
Cluster
Switches

HashinginNetworkedSystems

COS461:ComputerNetworks
Spring2011

MikeFreedman
h@p://www.cs.princeton.edu/courses/archive/spring11/cos461/

Hashing
HashfuncIon
FuncIonthatmapsalarge,possiblyvariablesized
datumintoasmalldatum,oNenasingleintegerthat
servestoindexanassociaIvearray
Inshort:mapsnbitdatumintokbuckets(k<<2n)
ProvidesIme&spacesavingdatastructureforlookup

Maingoals:
Lowcost
DeterminisIc
Uniformity(loadbalanced)

Todaysoutline
Usesofhashing
EqualcostmulIpathrouInginswitches
Networkloadbalancinginserverclusters
PerowstaIsIcsinswitches(QoS,IDS)
CachingincooperaIveCDNsandP2Plesharing
DataparIIoningindistributedstorageservices

Varioushashingstrategies
Modulohashing
Consistenthashing
BloomFilters

UsesofHashing

EqualcostmulIpathrouIng(ECMP)

ECMP
MulIpathrouIngstrategythatsplitstracover
mulIplepathsforloadbalancing

Whynotjustroundrobinpackets?
Reordering(leadtotripleduplicateACKinTCP?)
DierentRTTperpath(forTCPRTO)
DierentMTUsperpath

EqualcostmulIpathrouIng(ECMP)

PathselecIonviahashing
#buckets=#outgoinglinks
HashnetworkinformaIon(source/destIPaddrs)to
selectoutgoinglink:preservesowanity

Now:ECMPindatacenters

DatacenternetworksaremulIrootedtree
Goal:Supportfor100,000sofservers
RecallEthernetspanningtreeproblems:Noloops
L3rouIngandECMP:TakeadvantageofmulIplepaths

Networkloadbalancing
Goal:Splitrequestsevenlyoverkservers
Mapnewowstoanyserver
PacketsofexisIngowsconInuetousesameserver

3approaches
LoadbalancerterminatesTCP,opensownconnecIontoserver
VirtualIP/DedicatedIP(VIP/DIP)approaches
OneglobalfacingvirtualIPrepresentsallserversincluster
HashclientsnetworkinformaIon(sourceIP:port)
NATapproach:ReplacevirtualIPwithserversactualIP
DirectServerReturn(DSR)

LoadbalancingwithDSR
LB
Server
Cluster
Switches

ServersbindtobothvirtualanddedicatedIP
LoadbalancerjustreplacesdestMACaddr
ServerseesclientIP,respondsdirectly
PacketinreversedirecIondonotpassthroughloadbalancer
Greaterscalability,parIcularlyfortracwithassymmetric
bandwidth(e.g.,HTTPGETs)

10

Perowstateinswitches
SwitchesoNenneedtomaintainconnecIon
recordsorperowstate
Qualityofserviceforows
Flowbasedmeasurementandmonitoring
PayloadanalysisinIntrusionDetecIonSystems(IDSs)

Onpacketreceipt:
HashowinformaIon(packet5tuple)
Performlookupifpacketbelongstoknownow
Otherwise,possiblycreatenewowentry
ProbabilisIcmatch(falseposiIves)maybeokay

11

CooperaIveWebCDNs
TreeliketopologyofcooperaIvewebcaches
Checklocal
Ifmiss,checksiblings/parent

Oneapproach
InternetCacheProtocol(ICP)
UDPbasedlookup,shortImeout

public
Internet
Parent
webcache

AlternaIveapproach
Aprioriguessissiblings/childrenhavecontent
Nodessharehashtableofcachedcontentwithparent/siblings
ProbabilisIccheck(falseposiIves)okay,asactualICPlookupto
neighborcouldjustreturnfalse

12

HashtablesinP2Plesharing

Twolayernetwork(e.g.,Gnutella,Kazaa)
Ultrapeersaremorestable,notNATted,higherbandwidth
Leafnodesconnectwith1ormoreultrapeers

Ultrapeershandlecontentsearchers
Leafnodessendhashtableofcontenttoultrapeers
Searchrequestsoodedthroughultrapeernetwork
Whenultrapeergetsrequest,checkshashtablesofits
childrenformatch

13

DataparIIoning
Networkloadbalancing:Allmachinesareequal
DataparIIoning:Machinesstoredierentcontent
NonhashbasedsoluIon
DirectoryservermaintainsmappingfromO(entries)to
machines(e.g.,Networklesystem,GoogleFileSystem)
Nameddatacanbeplacedonanymachine

HashbasedsoluIon
NodesmaintainmappingsfromO(buckets)tomachines
Dataplacedonthemachinethatownsthenamesbucket

14

ExamplesofdataparIIoning
Akamai
1000clustersaroundInternet,each>=1servers
Hash(URLsdomain)tomaptooneserver
AkamaiDNSawareofhashfuncIon,returnsmachinethat
1. isingeographicallynearbycluster
2. managesparIcularcustomerdomain

Memcached(Facebook,Twi@er,)
Employkmachinesforinmemorykeyvaluecaching
Onread:
Checkmemcache
Ifmiss,readdatafromDB,writetomemcache
Onwrite:invalidatecache,writedatatoDB

15

HowAkamaiWorksAlreadyCached
cnn.com (content provider)

GET
index.
html
1

DNS root server

Akamai server

Akamai high-level
DNS server

2
7

Akamai low-level DNS

8
9
Enduser

10
GET /cnn.com/foo.jpg

Nearby
hash-chosen
Akamai
server

Cluster

16

HashingTechniques

17

BasicHashTechniques
Simpleapproachforuniformdata
IfdatadistributeduniformlyoverN,forN>>n
Hashfn=<data>modn
Failsgoalofuniformityifdatanotuniform

Nonuniformdata,variablelengthstrings
Typicallysplitstringsintoblocks
PerformrollingcomputaIonoverblocks
CRC32checksum
CryptographichashfuncIons(SHA1has64byteblocks)

18

ApplyingBasicHashing
ConsiderproblemofdataparIIon:
GivendocumentX,chooseoneofkserverstouse

Supposeweusemodulohashing
Numberservers1..k
PlaceXonserveri=(Xmodk)
Problem?Datamaynotbeuniformlydistributed
PlaceXonserveri=hash(X)modk
Problem?
Whathappensifaserverfailsorjoins(kk1)?
WhatisdierentclientshasdierentesImateofk?
Answer:Allentriesgetremappedtonewnodes!

19

ConsistentHashing
insert(key
lookup(key
1,value)
1)

key1=value

key1

key2

key3

ConsistenthashingparIIonskeyspaceamongnodes
Contactappropriatenodetolookup/storekey
Bluenodedeterminesrednodeisresponsibleforkey1
Bluenodesendslookuporinserttorednode

20

ConsistentHashing

0000
0010
URL
00011

0110

1010
URL
01002

1100

1110 1111

URL
10113

ParIIoningkeyspaceamongnodes
NodeschooserandomidenIers:

e.g.,hash(IP)

KeysrandomlydistributedinIDspace:

e.g.,hash(URL)

KeysassignedtonodenearestinIDspace
Spreadsownershipofkeysevenlyacrossnodes

21

ConsistentHashing
0

ConstrucIon
AssignChashbucketstorandompoints
onmod2ncircle;hashkeysize=n

14
12

Bucket

MapobjecttorandomposiIononcircle
Hashofobject=closestclockwisebucket

Desiredfeatures
Balanced:NobuckethasdisproporIonatenumberofobjects
Smoothness:AddiIon/removalofbucketdoesnotcause
movementamongexisIngbuckets(onlyimmediatebuckets)
Spreadandload:Smallsetofbucketsthatlienearobject

22

BloomFilters
DatastructureforprobabilisIcmembershiptesIng
Smallamountofspace,constantImeoperaIons
FalseposiIvespossible,nofalsenegaIves
UsefulinperownetworkstaIsIcs,sharinginformaIon
betweencooperaIvecaches,etc.

Basicideausinghashfnsandbitarray
UsekindependenthashfuncIonstomapitemtoarray
Ifallarrayelementsare1,itspresent.Otherwise,not

23

BloomFilters
Startwithanmbitarray,lledwith0s.
0

Toinsert,hasheachitemkImes.IfHi(x)=a,setArray[a]=1.
0

Tocheckifyisinset,checkarrayatHi(y).Allkvaluesmustbe1.
0

PossibletohaveafalseposiIve:allkvaluesare1,butyisnotinset.
0

24

Todaysoutline
Usesofhashing
EqualcostmulIpathrouInginswitches
Networkloadbalancinginserverclusters
PerowstaIsIcsinswitches(QoS,IDS)
CachingincooperaIveCDNsandP2Plesharing
DataparIIoningindistributedstorageservices

Varioushashingstrategies
Modulohashing
Consistenthashing
BloomFilters