Towards a Redundancy -Aware Network Stack for Data Centersmusa/uploads/hotnets_2016_talk.pdf · Towards a Redundancy -Aware Network Stack for Data Centers Ali Musa Iftikhar (Tufts)

TowardsaRedundancy-AwareNetworkStackforDataCenters

AliMusaIftikhar(Tufts)

Ihsan AyyubQazi(LUMS)

FahadR.Dogar(Tufts)

TheProblemofTailLatency inDataCenters!

2


Highfan-out

3


• Loadimbalance• Backgroundtasks• Failures,etc.

Highfan-out

Straggler

4


• Loadimbalance• Backgroundtasks• Failures,etc.

LongtaillatencyHighfan-out Stragglers+

Highfan-out

Straggler

5

Howtoavoidstragglers?

Reactively Proactively

6



PRO:lowoverhead

Hopper(SIGCOMM’15)C3(NSDI’15)

Sinbad(SIGCOMM’13)

CON:requiresstragglerdetection(slowandinaccurate)

7



PRO:lowoverhead

Dolly(NSDI’13)Lowlatencyvia

redundancy(CoNext’13)



PRO:fastandaccurate

CON:requiresdeterminingthresholdload(non-trivial)


8



PRO:lowoverhead

Can we achieve the benefits of both without their limitations?

Dolly(NSDI’13)Lowlatencyvia

redundancy(CoNext’13)



PRO:fastandaccurate

CON:requiresdeterminingthresholdload(non-trivial)


9

Overview• Duplicate-Aware Scheduling Framework

• Redundancy-Aware Network Stack

• Preliminary Results

Genericframework

NewnetworkstackforDC

10

Duplicate-awarescheduling

Replica1

Client

Replica2 11


Replica1

Client

Replica2

high

low

high

low

1. PriorityQueues

12


Replica1

Client

Replica2

high

low

high

low

request

1. PriorityQueues

13


Replica1

Client

Replica2

high

low

high

low

request

1. PriorityQueues

14


Replica1

Client

Replica2

high

low

high

low

P

request

B

1. PriorityQueues

15


Replica1

Client

Replica2

high

low

P

high

lowB

request

1. PriorityQueues

16


Replica1

Client

Replica2

high

low

high

lowB

request

1. PriorityQueues

17


Replica1

Client

Replica2

high

low

high

lowB

request

purge

1. PriorityQueues2. Purging

18

NeedforPriorityQueuing

high

lowbackup

primary

19

ØDuplicationhasanoverhead!

L


high

lowbackup

primary

üStrictprioritiesüWorkconservationüPreemption

20


LPropertiesrequired:


high

lowbackup

primary


21



PQ makes the overhead of duplication low. sJ


high

lowbackup

primary


22



PQ makes the overhead of duplication low. sJessential

ImportanceofPurging

ØStalerequestsblocknewrequests.

Lhigh

lowreq1req2

stale

23

ImportanceofPurging

high

lowreq1req2

stale


L

Purging makes the system more efficient! A J24

ImportanceofPurging

high

lowreq1req2

stale


L

Purging makes the system more efficient! A Joptimization 25

RealizingDuplicate-AwareSchedulingateverypotentialbottleneck resourceinaDC

26


Network

27


Compute

Network

28


Memory

Compute

Network

29


GFSHDFS BigTable

Memory

Compute

Filesystem/Database

Network

30


GFSHDFS BigTable

Storage

Memory

Compute

Filesystem/Database

Network

31

ateverypotentialbottleneck resourceinaDC

GFSHDFS BigTable

Memory

Compute

Storage

Network

Filesystem/Database

In-networkpurging

Prioritization

Purging+preemption

challenges

32

RealizingDuplicate-AwareScheduling

RedundancyAwareNetworkStack(RANS)

Application

Transport

Link

Network

Duplicate-Awareness

Pointtomultipoint

PriorityQueues

Physical Sameasbefore

Sameasbefore

ExpressiveInterface

Layer NewRole

+purging

+purging

33


Application

Transport

Link

Network

Duplicate-Awareness

Pointtomultipoint

PriorityQueues


Sameasbefore

ExpressiveInterface

Layer NewRole

+purging

Applicationsneedtobemodified.

challenge

+purging

34


Application

Transport

Link

Network

Duplicate-Awareness

Pointtomultipoint

PriorityQueues


Sameasbefore

ExpressiveInterface

Layer NewRole

+purging


Expressiveinterfaceallowsrichcommunicationb/wAppand

Transport.E.g.DAG

challenge

opportunity

+purging

35


Application

Transport

Link

Network

Duplicate-Awareness

Pointtomultipoint

PriorityQueues


Sameasbefore

ExpressiveInterface

Layer NewRole

+purging



Transport.E.g.DAG

challenge

opportunity

Hardtoimplementperpacketpurging.

challenge+purging

36


Application

Transport

Link

Network

Duplicate-Awareness

Pointtomultipoint

PriorityQueues


Sameasbefore

ExpressiveInterface

Layer NewRole

+purging



Transport.E.g.DAG

challenge

opportunity


challenge

AddssupportforexistingPQsinDCswitches.

opportunity

+purging

37


Application

Transport

Link

Network

Duplicate-Awareness

Pointtomultipoint

PriorityQueues


Sameasbefore

ExpressiveInterface

Layer NewRole

+purging



Transport.E.g.DAG

challenge

opportunity


challenge

AddssupportforexistingPQsinDCswitches.

opportunity

+purging

38

e.g.Improvedfaulttolerance

üMultipath

üMulti-destination

RANSTransport:PointtoMulti-point

ØEnables:Richtransport

Sender1(replica1)

Receiver(client)

Sender2(replica2)

39

RANSTransport:ByteAggregation

Sender1(replica1)

Receiver(client)

Sender2(replica2)

ØOpportunity:Receiverdriventransport

Response

e.g.Moreefficientcongestioncontrol(2xormore)

üTwoormoreresponsestreams

üAggregatebytesatreceiverside

40

RANSTransport:PriorityAssignment

Sender1(replica1)

Receiver(client)

Sender2(replica2)

ØDynamicreplicaassignment

Response+Feedback

e.g.Improvedreplicaassignment

üFinegrainedmonitoringofcongestionwindow

üDynamicallyreprioritizeflows

üFeedbacktoApplication

41

Overview• Duplicate-Aware Scheduling Framework

• Redundancy-Aware Network Stack

• Preliminary Results

42

PreliminaryEvaluation:ns-2setupdetailsØ Storagescenario

Client

10servers

Replica1

Replica2

43

PreliminaryEvaluation:ns-2setupdetailsØ Storagescenario

Client

10servers

Replica1

Replica2

bottlenecks

44

PreliminaryEvaluation:ns-2setupdetails

TrafficDetails

Totalrequests 20K

Arrivalprocess Poisson

Server&replicaselection Uniformlyrandom

Ø Storagescenario

Client

10servers

Replica1

Replica2

bottlenecks

45

PreliminaryEvaluation:ns-2setupdetails

TrafficDetails

Totalrequests 20K

Arrivalprocess Poisson

Server&replicaselection Uniformlyrandom

Ø Storagescenario

Client

10servers

Replica1

Replica2

bottlenecks

The only source of stragglers is load imbalance.

46

Noduplicates(baseline)

2-copies(proactivew/oPQ)

+PQs

+Purging

+ByteAggregation(RANS)

Averagerequestcompletiontimeof:

Load(%)

Requ

estcom

pletiontim

e(s)

47



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

48



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

49



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

50



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

51



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

~2X

52



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

~2X

Expecting more gains even at lower loads with additional straggler sources.

53



+PQs

+Purging



Load(%)

Requ

estcom

pletiontim

e(s)

50-80% improvement over the baseline across all loads.

Expecting more gains even at lower loads with additional straggler sources.

~2X

54

Summary&Futurework

• TheIssueofStragglers

• Duplicate-AwareSchedulingFramework

• RANS

• ImplementinginHDFSandCassandra

Simpleyetchallengingsolution

Afirststeptowardsaduplicate-awarenetwork

55

RANS:FeedbackandDiscussion

• AliMusaIftikhar([email protected])

• FahadR.Dogar ([email protected])

• Ihsan A.Qazi ([email protected])Transport

Link

Network

PointtomultipointByteaggregationPriorityassignment

PriorityQueues


Sameasbefore

ExpressiveInterface

Application Duplicate-Awareness

Layer NewRole

+purging

+purging

56

Possiblequestions– backupslide• Preemptionoverhead

• Notreallyanissueinthenetworkbecausepacketsaresmall.

• Packetpurging• PFC(backpressure,buildqueuesattheendhostsandpurgethem)

• Droptheentireduplicatequeue(easierthanper-packetdrops)

• Recenttrendtowardsprogrammableswitches

• GainswithPQ• Moregainswithfailuresasstragglers(primaryundergoesafailure)

• Alsomorebenefitswithdifferentresources

• Duplicationoverheadatclient• Clientisusuallynotthebottleneck

• Non-Idempotentrequests• Wearetargetingtheclassofappswhichhaveflexibleendpointsandrequireatleastoncesemantics

• Replicatingonlysmallpacketsandprioritizingthem• Onlybeneficialwithbursty smallflows• HDFShaveatypicalchunksizeb/w64MB-128MB

• Quorumsystems• RANScomplementssuchsystems,theycanusethistechniqueandsendKoutofNrequestsathighprio whileN-Kasbackups

• Can’tjustimplementattheappandgetthesamebenefits?• Networkcouldbeabottleneck• Finegrainedcontrol,muchmorecontrol

• Rootcausesofperformanceimprovement• PQavoidsoverheads• Nowwecaneasilygetthebenefitsofduplicationslikeaggregationetc.

• Purgingwillalsoattimespurgeprimarymakingthesystemmoreefficient.

57

Foodforthought

DCPrimary DCFailover

InterDCDuplicate-AwareScheduling

e.g.Google’sGeo-DistributedDatabase“Spanner”(OSDI’12)

58

Foodforthought

DCPrimary DCFailover

InterDCDuplicate-AwareScheduling

e.g.Google’sGeo-DistributedDatabase“Spanner”(OSDI’12)

Prefetch

SearchsuggestionsSpellcheck

• Searchenginesdropspellcheck,suggestions,etc.athighloads.

• Canbenefitfromduplicate-awarescheduling.

59

WhenRANSworksbest?

• Applicationfanout ishighandstragglersarefrequent.• End-pointsareflexibleand“atleastonce”semanticsaresufficient.• Clientisnotthebottleneck.• Requestsizesaresmall(orpreemptionoverheadisminimal).

60

Towards a Redundancy -Aware Network Stack for Data Centersmusa/uploads/hotnets_2016_talk.pdf · Towards a Redundancy -Aware Network Stack for Data Centers Ali Musa Iftikhar (Tufts)

Documents