granular compu,ng and network intensive applica,ons...

26
Granular Compu,ng and Network Intensive Applica,ons: Friends or Foes? Arjun Singhvi, Sujata Banerjee, Yotam Harchol, Aditya Akella, Mark Peek, Pontus Rydin

Upload: others

Post on 19-Jul-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

GranularCompu,ngandNetworkIntensiveApplica,ons:FriendsorFoes?

ArjunSinghvi,SujataBanerjee,YotamHarchol,AdityaAkella,MarkPeek,PontusRydin

Page 2: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Evolu,onofCompu,ngServices

BareMetal

GranularServerlessCompu,ng

Decreaseuserinvolvementinmanagementtasks

Allowmoreuserfocusonapplica,onlogic

VirtualMachines Containers

Page 3: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ProgrammingParadigmShiL

Tradi,onal

MonolithicApplica,on

Serverless

Collec,onofFunc,ons

Page 4: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ServerlessCompu,ngControlFlow

Upload your function to the

serverless computing platform

and register for event-based

triggers

Execution is triggered when an

event occurs

Platform runs your function

ONLY when triggered

Pay just for the computation time

Page 5: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ServerlessCompu,ngBuildingBlocks

Func,onScheduler

f()KeyValueStore

BlobStore

ServerlessFunc,onsare

Stateless!

Page 6: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ServerlessCompu,ngPlaSorms

UW-Madison

Amazon Google MicrosoL IBM

VMware

Page 7: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ResearchonServerlessCompu,ng

•  VideoProcessingApplica,ons– ExCamera(NSDI’17)

•  DistributedCompu,ngApplica,ons– PyWren(SoCC’17)

•  NetworkIntensiveApplica,ons->NetworkFunc,ons– OurWork(HotNets‘17)

Page 8: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

NetworkFunc,ons

•  Examineandperformstatefulac,onsonpackets/flows

•  Ensuresecurity,improveperformanceandprovideothernetwork-relatedfunc,onality

•  Lieinthecri,calpathbetweensourceanddes,na,on

•  Shouldbecapableofhandlingpacketburstsandfailures

CacheIDSFirewall LoadBalancer

Page 9: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

NetworkFunc,onsStateTaxonomy

StatecreatedorupdatedbyanNFappliestoeitherasinglefloworacollec,onofflows

Connec,on

Connec,on

TcpAnalyzer

HbpAnalyzer

TcpAnalyzer

HbpAnalyzer

Per-flowstate

ConnCount

Cross-flowstate

Page 10: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Intrusiondetec,on

system(IDS)

Cachingproxy

NetworkFunc,onsVirtualiza,on

•  Leveragecloudcompu,ngandSDNinTelcoinfrastructures– U,lizecommodityhardware– Reducecosts– DynamicallyallocateNFinstances–  Introducenewservicesfast

Hypervisor

Page 11: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

WhyNetworkFunc,ons?

•  NetworkFunc,onsareademandingclassofapplica,ons

•  TrendofpushingNetworkFunc,onstothecloud– APLOMB(SIGCOMM‘12),EMBARK(NSDI‘16),AT&TDomain2.0

•  Inthefuture,ifserverlesscompu,ngbecomesthede-factocloudstandard

Page 12: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Whyisitchallengingtorunnetworkfunc,onsatopserverless

compu,ngplaSorms?

Page 13: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Short-lived

Stateless

Nosupportforefficientchaining

NoQoS

guarantees

Long-lived

Stateful

Needsefficientchaining

NeedQoSguarantees

ServerlessCompu,ng NetworkFunc,ons

Foes?

Page 14: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Whatopportuni,esdoserverlesscompu,ngplaSormsprovideto

runnetworkfunc,ons?

Page 15: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

•  NetworkFunc,onsVision–Micro-servicebasedarchitecturethatscalesdynamically–  Beberresourceu,liza,on–  Scalecomponentsindependently–  Reuseacrosschains–  Reducethecosts

•  ServerlessCompu,ngprovidesthebuildingblocks– Hasthecapabilitytoscalebasedonload–  Encouragesdeveloperstodesigntheirapplica,onsintermsofmicro-services

–  Reducesthecosts

Friends?

Page 16: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

40Gbps

Firewall

HourlyCost

$8

60¢

Page 17: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ServerlessPacketProcessingDesign

•  Whatisthegranularityatwhichwewanttolaunchlambdas?

LambdaScheduler

•  NotviableCostPerformance

Per-packet

Page 18: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

ServerlessPacketProcessingDesign

•  Whatisthegranularityatwhichwewanttolaunchlambdas?Per-flow

LambdaScheduler

Packets flow

Processes Packets

Coordinator

Gateways

RemoteDatastore

Page 19: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Per-flowVs.Per-packet•  Pros–

– Costeffec,ve– Statemanagementeasier

•  Cons–– Needsaddi,onalinfrastructuretocoordinatelaunchingofnewlambdas

– Cannotna,velyleveragewhatserverlessplaSormsprovides

Page 20: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

AWSLambdaBenchmarkingStudy

•  WithinanAWSregion•  EC2VMsactasTCPtrafficsourcesandsinks•  ThreeClick-basedNFs–

– PacketCounter– Firewall–  IntrusionDetec,onSystem

Page 21: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

AWSLambda–NetworkProvisioning

128M 256M 512M 1GB 1.5GLambda Memory Size

400420440460480500520540560

Band

widt

h (M

bps)

OutboundInbound

Bidirec,onalBandwidthofaround500Mbps

VM

Page 22: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

1 2 3 4 5 6 7 8Simultaneous Request Count

500

1000

1500

2000

2500Th

roug

hput

(Mbp

s)

•  Bandwidthdoesnotscalelinearly•  Noperlambdarequestnetworkguarantees

AWSLambda–NetworkProvisioning

Page 23: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

AWSLambda–ChainingOverhead

Highchainingoverheadobserved

1 2 3 4Lambda Chain Length

0

100

200

300

400

500

600

Ove

rhea

d (m

s)

Page 24: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

AWSLambda–NFFramework

Modestresults;butgiveusahopeastowhatfutureserverlessplaSormsmayenable

SourceGateway

SinkGateway

Page 25: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

Conclusion

•  ServerlessCompu,ngPlaSormsarenotperfectyet–arecon,nuouslyevolving!

•  Packetprocessingapplica,onsarenotanaturalfit;need–– Efficientin-networktriggers– Efficientchainingsupport– Bandwidthguarantees

Page 26: Granular Compu,ng and Network Intensive Applica,ons ...netseminar.stanford.edu/seminars/03_15_18.pdf · Serverless Compu,ng Control Flow Upload your function to the serverless computing

OpenQues,ons

•  Scalableremotestorageservicethatmeetsourrequirements?

•  Beberfaulttolerancesupport?•  Fastercontrolplanedecisions?•  Andmanymore…Howtoeffec:velysupportdemandingstateful

applica:onsatopserverlesspla@orms?