Scaling MongoDB with Docker and cgroups

Marco Bonezzi Technical Services Engineer, MongoDB

marco@mongodb.com

@marcobonezzi

2

Agenda

•  MongoDBandWiredTiger

•  WhyshouldyouuseDockerwithMongoDB?

•  ScalingMongoDBwithDocker

•  Results

PS

S

config

SS

PS

S

SHARD2

SHARD1

mongos

MongoDB and WiredTiger

5

MongoDB3.0andWiredTiger

MongoDBStorageEngineAPI+WiredTigerstorageengineMainfeatures•  Document-levelconcurrencycontrol•  Durability•  Compression•  EncrypHonatrest(3.2)

6

WiredTigerincreasedscalability

•  MulH-corescaling

•  Lock-freealgorithms(hazard-pointers)

•  Noin-placeupdate(skiplists)

•  Eliminateblockingwithconcurrencycontrol

7

WiredTigerincreasedscalability

•  Higherthroughputonwriteandupdates(skiplists)

8

Bestprac>cesforWiredTiger

•  XFSfilesystem

•  #concurrentac+veopera+ons•  Bestthroughputwhen<=#CPU

•  MemoryandWiredTigercache

•  storage.wiredTiger.engineConfig.cacheSizeGB

9

MongoDBMemoryandWiredTigercache

cacheSizeGB=max(1GB,0.5*memory)

WiredTigercache: Usedfordataandindexesonly

10

MongoDBMemoryandWiredTigercache

WiredTigercache: Usedtostorethecurrentworkingset

cacheSizeGB=max(1GB,0.5*memory)

11

MongoDBMemoryandWiredTigercache

MongoDBMemory:mongodprocess:connecHons,aggregaHons,mapReduce,etc

cacheSizeGB=max(1GB,0.5*memory)

Totalmemory=mem(mongod)+cacheSizeGB

12

ContainerizingMongoDB

•  MongoDBdockerfile(orcreateyourown)hPps://github.com/sisteming/mongodocker/[mongodb.docker]

•  Genericmongod.confconfiguraHonfile•  Automateyourdeploymentforscaling:

•  #mongods• Portsandnetworkinterfaces•  SingleormulHplehosts• CPUormemorylimits

13

Summary

•  MongoDBwithWiredTiger:•  Higherthroughputwithincreasedscalability•  AddiHonalfeatures:compression,encrypHon•  WiredTigercachememory

•  MongoDBonDocker:•  Genericimage•  Automatetoscale

Why should you use Docker for MongoDB?

15

DockerAccordingtodocker.com:

HowcanthisbeusefulformyMongoDBdeployments?

(That’swhyyouareheretoday)

16

WhyshouldyouuseDockerforMongoDB?

Build•  anddefineyour

containerbasedonOS,mongodbreleaseorconfiguraHon

•  Buildonce,deployeverywhere

17

WhyshouldyouuseDockerforMongoDB?

Ship•  DeploythesameimageinDev,Pre-prod(or

evenProd!)withthesamesetup•  MongoDBstandalone,ReplicaSetsorcluster

withmulHpleshardsfromthesameimage

18

WhyshouldyouuseDockerforMongoDB?

Run•  YourmongoDBprocessesisolatedfromthe

restofprocessesonthesystem•  Isolatetheresourcesforco-locatedmongodb

processes

19

VS

20

Resourcemanagement

CgroupsImplementaHononDocker

21

Resourcemanagement

CgroupsImplementaHononDocker

--cpu-shares--cpuset-cpus

--memory

--blkio-weight

--net

22

Summary

UsingDockerforMongoDBisusefulformanyreasons:

•  Buildonce,deployeverywhere•  Environmentportability•  MongodprocessisolaHon•  Fasterdeploymentsofcomplexclusters•  Resourcemanagementwithcgroups

Scaling MongoDB with Docker

config

SS

PS

S

PS

S

SHARD2

SHARD1

mongos

24

UsecasesofMongoDBandDockerMicrosharding

•  MulHpleco-locatedmongodinstances

•  Eachmongod=shard•  AggregaHons

Mul>pleinstances•  Controlco-locatedmongod

processesonsameserver•  Differentinstances/uses

25

Memorylimit

MEM=($TOTAL_MEMORY-2048)/$NUM_MONGODWiredTigercache=$MEM/2

26

CPUlimit

NUMBER=from0toMAX_CPU_CORES(i.e.with32cores->--cpuset=0,16)

27

Docker+cgroups:Memorymapping

WiredTigercachesize:definedforeachshard50%MongoDBmemory–50%WiredTigercache

28

MongoDBonDocker+cgroups:CPUmapping

Eachshardmappedto1(ormore)core(using--cpuset-cpus)

core0 core1 core2 core3

core4 core5 core6 core7

core8 core9 core10 core11

core12 core13 core14 core15

29

ResourceusagewithDocker•  Understandingresourceusage:•  ContainerstatsavailablethroughDockerremoteAPI:

GET/containers/(id)/statsecho-e"GET/containers/$id/statsHTTP/1.0\r\n"|

nc-U/var/run/docker.sock

30

Summary

•  UsesofscalingMongoDBwithDocker

•  LimiHngCPUonyourcontainer

•  DefiningmemoryandWiredTigercachelimits

•  ResourceusagewithDocker

Results

32

TargetandscenariosScalabilitybymicroshardingMongoDBwith

WiredTigerusingdocker(andcgroups)

•  Maintestscenarios:– ShardedclusterwithP/S/SReplicaSet

•  From2upto16shards– Eachscenario:

• numberofthreads:8,16,32• Write/ReadraHo:95/5• memoryandWiredTigercache

33

Workload

POCDriver (https://github.com/johnlpage/POCDriver) • Easy to use

–  java -jar POCDriver.jar -d 610 -c mongoDB://node1:37019 -o test.log -i 95 -t 10 -k 5 –e

• Automatic collection sharding with even distribution

34

Workloadresultssummary

ScenarioAMEM=2GB

WiredTigercache=1GB

ScenarioBMEM=32GB/$NUM_SHARDS

WiredTigercache=$MEM/2(GB)

35

TotalInserts

36

Inserts/second

37

Workloadresultssummary

3xc3x4largeAWSinstances:16cores32GBram

SSDvolumes

0

10

20

30

40

50

0 2 4 6 8 10 12 14 16 18

Millions

TotalInsertsvs#shards

TotalInserts

0

10

20

30

40

50

0 2 4 6 8 10 12 14 16 18

Millions

TotalInsertsvs#shards

TotalInserts

ScenarioAMEM=2GB

WiredTigercache=1GB

ScenarioBMEM=32GB/$NUM_SHARDS

WiredTigercache=$MEM/2(GB)

Summary

39

Keys for scaling MongoDB and WiredTiger with Docker

•  Startfromagenericimage,automatetoscale

•  Setlimitstoallyourcontainers

•  Define–WiredTigerCacheSizeGBoneachmongod

•  EsHmateanddefineresourcedistribuHonforyourarchitecture

•  Deployandtestwithyourownworkload

Questions?

Thank You!

Marco Bonezzi

@marcobonezzi

marco@mongodb.com

#MDBDaysmongodb.com

Getyourtechnicalques+onsansweredBenjaminBri3enlounge(3rdfloor),10:00-17:00Byappointmentonly–registerinperson