Predictive Elastic Database Systems - · PDF fileE-Store Elastic DBMS Architecture 5 Live Migration ... Machine Capacity Effective Capacity 13. Options 1. ... Dynamic Programming Algorithm

Predictive Elastic Database Systems

RebeccaTaft–[email protected]

1

ModernOLTPApplications

Large Scale Cloud-Based Performance is Critical

2

Challengestotransactionperformance:skewandworkloadvariation

3

DatabaseElasticityE-Store–manageskewandreacttovariation

P-Store–predictivemodelingfortimevariation

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

4

UniformWorkload,IncreasingLoad

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

4

UniformWorkload,IncreasingLoad

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3 Part4 Part54

UniformWorkload,IncreasingLoad

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3 Part4 Part54

UniformWorkload,IncreasingLoad

TransacHon

s

0

2.5

5

7.5

10

Part1 Part2 Part3 Part4 Part5

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

4

UniformWorkload,IncreasingLoad

TransacHon

s

0

2.5

5

7.5

10

Part1 Part2 Part3 Part4 Part5

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3

TransactionArrivalRate

4

UniformWorkload,IncreasingLoad SkewedWorkload

TransacHon

s

0

2.5

5

7.5

10

Part1 Part2 Part3 Part4 Part5

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3

TransactionArrivalRate

4

UniformWorkload,IncreasingLoad SkewedWorkload

TransacHon

s

0

3

5

8

10

Part1 Part2 Part3

TransacHon

s

0

2.5

5

7.5

10

Part1 Part2 Part3 Part4 Part5

E-Store: Elastic Scaling to Adapt to Workload

TransactionArrivalRate TransactionArrivalRate

4

UniformWorkload,IncreasingLoad SkewedWorkload

E-Store Elastic DBMS Architecture

5

LiveMigration

ElasticController

PlannerLoadMonitor

SharedNothingDBMS(e.g.H-Store)

ACaseStudy:B2WDigital

6

3-Day Online Retail Database Workload

7

3-Day Online Retail Database Workload

7

3-Day Online Retail Database Workload

7

3-Day Online Retail Database Workload

7

3-Day Online Retail Database Workload

7

Elastic Scaling Adapts to Workload

8

Reactive Scaling Causes Latency Spikes

8

P-StoreA Predictive Elastic DBMS

9

P-Store Architecture

10

LiveMigration

PredictiveController

Predictor PlannerLoadMonitor

SharedNothingDBMS(e.g.H-Store)

Ideal Capacity Actual Servers Allocated

DemandMachineCapacity

11

DemandMachineCapacity

12

?

DemandMachineCapacity

13

Options1. Addmachine(s)

DemandMachineCapacity

13

Options1. Addmachine(s)

DemandMachineCapacity

13

Options1. Addmachine(s)

DemandMachineCapacity

13

Options1. Addmachine(s)2. Removemachine(s)

DemandMachineCapacity

13

Options1. Addmachine(s)2. Removemachine(s)3. Nochange

DemandMachineCapacity

13

Options1. Addmachine(s)2. Removemachine(s)3. Nochange

DemandMachineCapacity

13

Options1. Addmachine(s)2. Removemachine(s)3. Nochange

DemandMachineCapacity

EffectiveCapacity

13

Options1. Addmachine(s)2. Removemachine(s)3. Nochange

Challenges1. TimetoReconfigure2. EffectiveCapacity3. Cost

DemandMachineCapacity

EffectiveCapacity

13

Time to Reconfigure

• Ratecontrol:smallchunksofdata,spacedapart• ∃sometimeD–minimumtimetomoveentiredatabasew/noperformanceimpact

%ofD

ata

0

50

100

Nodes

1 2 3 4

%ofD

ata

0

50

100

Nodes

1 2 3

→

Time = ¼ * D 14

Effective Capacity%of

TransacHon

s

0255075100

Nodes

1 2 3 4 5

%of

TransacHon

s

0255075100

Nodes

1 2

=

15

• Whattransactionratecanthesystemsupport?• Maxratepermachine:Q

Cost

Cost = 7

Nod

esAllocated

0

1

3

4

5

Time

1 2 2.1 3 3.1 4 4

DemandMachineCapacity

EffectiveCapacity

4machines att=9

Goal:findpathwithminimalcostsuchthateff.capacity>demand

Costof{pathtoAmachinesendingattimet}

= Costof{pathtoBmachinesendingattimet–TimeB→A}+

CostB→A 17

Dynamic Programming Algorithm for Planning Reconfigurations• Complexity∝#timestepsandmaxnumberofmachinesneeded• Inpractice:runtime<1second• Greedyalternativesdon’twork

18

DemandMachineCapacity

EffectiveCapacity

19

DemandMachineCapacity

EffectiveCapacity

OptimalSolution TotalCost:30

19

Time Series Prediction

• Diurnal,periodicworkload,somevariationdaytoday• Predictionmustcompleteinseconds-to-minutes• WeuseSparsePeriodicAuto-Regression(SPAR)–Chenetal.NSDI‘08

20

Evaluation

• Canwereduceresourceusage?• Canwepreventlatencyspikes?

21

Experiments

• 3DaysofB2WworkloadrunonH-Store• Clusterof10servers,6partitionsperserver• About1GBofshoppingcartandcheckoutdata• Trackmachinesallocated,throughput,latency,reconfigurationtimeperiods

22

Results – Static, Peak Provisioning Machines Used: 10

23

Results – Static, Average Provisioning Machines Used: 4

24

25

Results – Reactive ScalingAvg. Machines Used: 4.02

Results – P-Store with SPAR Avg. Machines Used: 5.05

26

Results – P-Store with Exact Provisioning Avg. Machines Used: 4.89

27

Results Comparison: CDF of Top 1% of Latency

28

Summary: P-Store Evaluation

• Canwereduceresourceusage?• Saves50%ofcomputingresourcescomparedtostaticallocation

• Canwepreventlatencyspikes?• Superiorperformancecomparedtoreactiveapproach

• Onarealworkload,P-Storereducesresourceusagewhilekeepinglatencywithinapplicationrequirements

29

Summary

❖Realdatabaseworkloadsareskewedandvaryovertime❖Elasticityenablesmanagementofskewandadaptationtoloadchanges

❖Predictivescalingimprovesperformanceduringloadchangescomparedtoreactivescaling

[email protected]

30