outline - university of washingtonfaculty.washington.edu/wlloyd/slides/prelim.pdf8/7/2012 1 phd...

8/7/2012

1

PHD Dissertation Proposal Defense

Wes J. LloydAugust 7, 2012

Colorado State University, Fort Collins, Colorado USA

Outline

Research ProblemResearch Problem

Challenges

Approaches & Gaps

Research Goals

Research Questions & ExperimentsResearch Questions & Experiments

Research Contributions

Preliminary Results

August 7, 2012 2Wes J. Lloyd PHD Dissertation Proposal Defense

Traditional Application Deployment

Data Business Logging App Data

Spatial DB

rDBMS

Business

Services

Services

Logging

Tracking DB

ppServer

Apache Tomcat

Research Problem 3

DODB / NOSQL

Services

ServicesObject Store

Single Server

Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

Infrastructure as a Service (IaaS) Cloud ComputingServer partitioning of multi‐core serversServer partitioning of multi‐core servers

Hardware virtualization

Service isolation

Resource elasticity

August 7, 2012 Research Problem 4Wes J. Lloyd PHD Dissertation Proposal Defense

8/7/2012

2

Research Problem 5Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

Problem Statement

Autonomic deployment of multi‐tierAutonomic deployment of multi‐tier applications to IaaS cloudsComponent composition Collocation and interference of components

Scaling infrastructure to meet demandScaling infrastructure to meet demand

August 7, 2012 Research Problem 6Wes J. Lloyd PHD Dissertation Proposal Defense

7Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

Application Component CompositionApplication

ComponentsVirtual Machine (VM)

App Server

ComponentDeployment

Virtual Machine (VM) Images

rDBMS r/o

File Server

Log Server

Load BalancerImage 2

rDBMS write

Image 1

App Server

File Server

Log Server

rDBMS writerDBMS r/o

Load Balancer

Problems & Challenges 8

Application “Stack”

PERFORMANCE

. . . Image n

Load BalancerDist. cache


8/7/2012

3

Bell’s Numbern = number

ofapplication n k

Model

ComponentDeployment

applicationcomponents VM deployments

Database

File Server

Log Server

config 1

M D

F L

config 2

M F

L

config n

M L D

1 VM : 1..n components

4 15

5 52

6 203

7 877

8 4,140

D

Problems & Challenges 9

Application “Stack”

# of Configurations

. . .k= # of

possibleconfigs

F 9 21,147

n . . .


Provisioning VariationRequest(s) to launch VMs VMs Share PM

CPU / Disk / NetworkVM

Physical Host Physical HostPhysical Host

VM VM

VM

AmbiguousMapping

VM

VM VM

VM

VM

VM

VM

VM

VM VM

VM VM VM

VM

Problems & Challenges10

Physical Host Physical Host Physical Host

VMs ReservePM Memory BlocksPERFORMANCE


Service Requests

Infrastructure Management• Scale Services

• Tune Application

ApplicationServers

Load Balancer

Load Balancer

distributed cache

• Tune Application Parameters

• Tune Virtualization Parameters

noSQL data storesrDBMS

Problems & Challenges 11Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012 12Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

8/7/2012

4

Virtual Machine (VM) Placement as “Bin Packing Problem”

Bins= physical machines (PMs)Bins physical machines (PMs) Items= virtual machines (VMs) Dimensions CPU time VM RAM, hard disk size, # cores Disk read/write throughput Disk read/write throughput Network read/write throughput

PM capacities vary dynamically VM resource utilization varies

Approaches & Gaps 13Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

NP-Hard

Related Work

Multivariate performance modelsMultivariate performance models Regression modelsMachine learning

Feedback loop controlHybrid approachesFormal approaches Integer linear programming Case based reasoning


Gaps in Related Work Existing approaches do not consider VM image composition VM image composition Complementary component placements Interference among componentsMinimization of resources (# VMs) Load balancing of physical resources

Performance models ignore Performance models ignore Disk I/O Network I/O VM and component location


Why Gaps Exist

Public cloudsPublic clouds Research is cost prohibitive

Users concerned with performance not in control

Private clouds: systems still evolving

Performance models (large problem space)Performance models (large problem space)

Virtualization misunderstood or overlooked


8/7/2012

5


Research Goals

RG1: Support VM component composition

RG2: Support virtual infrastructure management Determine and execute VM placement

Scale infrastructure for application demand

August 7, 2012 Research Goals 18Wes J. Lloyd PHD Dissertation Proposal Defense

Performance Objectives

Primary: Maximize application throughputPrimary: Maximize application throughput

Secondary: Minimize resource cost (# of VMs)

Minimize modeling time

Support high responsiveness to change in application demandapplication demand

August 7, 2012 Research Goals 19Wes J. Lloyd PHD Dissertation Proposal Defense 20Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

8/7/2012

6

Methodology Evaluation

CSIP: USDA‐NRCS platform for model servicespModels as multi‐tier application surrogates RUSLE2 – Soil erosion modelWEPS – Wind Erosion Prediction System Hydrology models: SWAT, AgES Other models: STIR SCI Other models: STIR, SCI…

Eucalyptus IaaS cloud(s) Amazon EC2 compatible XEN & KVM hypervisors

Research Questions & Methodology 21Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

ComponentComposition

InfrastructureManagement

RQ1RQ3

RQ4

August 7, 2012 Wes J. Lloyd PHD Dissertation Proposal Defense Research Questions & Methodology 22

RQ2RQ5

RQ1: Which independent variables best help model application performance (throughput) to guide autonomic component composition?

Total (all VMs) resource utilizationTotal (all VMs) resource utilization CPU time, disk I/O, network I/O, …

Individual VM and PM resource utilization

Component and VM location

VM Configuration: number of cores RAMVM Configuration: number of cores, RAM, hypervisor type (KVM, XEN...)


RQ1: Which independent variables best help model application performance (throughput) to guide autonomic component composition?

Total (all VMs) resource utilizationMethodologyExploratory performance modelingTotal (all VMs) resource utilization

CPU time, disk I/O, network I/O, …

Individual VM/PM resource utilization

Component / VM location

VM Configuration: number of cores RAM

Exploratory performance modeling

•Investigate independent variables•Investigate modeling techniques

• Multiple linear regression (MLR)• Artificial neural networks (ANNs)VM Configuration: number of cores, RAM,

hypervisor type (KVM, XEN...)


• Artificial neural networks (ANNs)• Others

8/7/2012

7

RQ2: Can component resource classifications and behavioral rules predict performance of component compositions?

Support simplification of the search space

n k

4 15

5 52

6 203


Support applications with large # of components Bell’s number

7 877

8 4,140

9 21,147

n . . .


RQ2: Can component resource classifications and behavioral rules predict performance of component compositions?

Support simplification of the search spaceMethodologyInvestigate autonomic component composition

n k

4 15

5 52

6 203


Support applications with large # of components e.g. Bell’s number:

Investigate autonomic component composition approach(es)

•Performance modeling•Heuristics to classify

• Component resource utilization 7 877

8 4,140

9 21,147

n . . .


• Component resource utilization• Component dependencies

Evaluation Metrics: Component Composition

C i i fComposition performance Average throughput of configurations

Resource packing density # components/# VMs for compositions

D i ti dDerivation speed Average wall clock time to produce compositions


RQ3: Does performance of component compositions change when scaled up?

Single provisioned application VMs

Multiply provisioned application VMs

Investigate collocation of new VMs Intelligent vs. ad‐hoc placement

Load balance physical resources

Research Questions & Experiments 28Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

8/7/2012

8

RQ3: Do performance rankings of component compositions change when scaled up?

MethodologySingle provisioned application VMs

Multiply provisioned application VMs

Investigate collocation of new VMs Intelligent vs. ad‐hoc placement

MethodologyScale up compositions and benchmark performance

•Investigate impact of VM placement for infrastructure scaling

Load balance physical resources


infrastructure scaling

RQ4: How rapidly can VMs be launched in response to application demand?

Determine upper bound of VM launch speedDetermine upper bound of VM launch speed

Devise workarounds to improve performance VM prelaunch and suspension Reserve RAM, other resources multiplexed

Enforced caching of VM data on PMs

Reassign duties of existing VMs

Others ?


RQ4: How rapidly can VMs be launched in response to application demand?

Determine upper bound of VM launch speedDetermine upper bound of VM launch speed

Devise workarounds to improve performance VM prelaunch and suspension

Reserve RAM, other resources multiplexed

Enforced caching of VM data on PMs

MethodologyBenchmark VM launch performance and investigate potential improvements

Reassign duties of existing VMs

Others ?


RQ5: Which independent variables best support application performance modeling for autonomic infrastructure management?

Virtual infrastructureVirtual infrastructure Number of VMs (1 to n) per application VM

VM RAM, # cores

VM location data

Application specific parameterspp p p Number of worker threads

Number of database connections

Number of app server concurrent connectionsResearch Questions & Methodology 32Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

8/7/2012

9

RQ5: Which independent variables best support application performance modeling for autonomic infrastructure management?

When resources are scaled ↑↓Methodology ↑↓ Virtual infrastructure Number of VMs (1 to n) per application VM VM RAM allocation VM core allocation Location of VMs

l f

Explore autonomic infrastructure management approach(es)

•Performance model based•Feedback control Application specific parameters Number of worker threads Number of database connections Number of app server concurrent connections


•Hybrid approach

Evaluation Metrics:Infrastructure Management

Percentage of service requests completedPercentage of service requests completed

ResponsivenessMax supported load acceleration without dropping requests

Adaptation timeAdaptation time Time window with dropped requests

Failure recovery time



Expected Contributions (1/2)

Novel intelligent approaches for IaaS cloudNovel, intelligent approaches for IaaS cloudApplication deployment

Infrastructure management

Move IaaS infrastructure management beyond simple management of VM poolsbeyond simple management of VM pools

August 7, 2012 Contributions 36Wes J. Lloyd PHD Dissertation Proposal Defense

8/7/2012

10

Expected Contributions (2/2)

Autonomic component composition (RQ1, RQ2)

Autonomic infrastructure management (RQ3, RQ4, RQ5)

Improve application performance modeling For component composition (RQ1) New independent variables (RQ1) Heuristics (RQ2) For infrastructure management (RQ5)

Support load balancing of physical resourcesContributions 37Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

Non‐goals

Support for stochastic applicationsSupport for stochastic applications Only applications with stable resource utilization characteristics supported

External interference From non‐application VMs

Hot‐spot detection

Contributions 38Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012


Component CompositionsSC2

M DF

L

SC6

M D F L

SC11

M F D L

Rusle 2 Model Fastest Compositions

F

Service Isolation Overhead

Preliminary Results 40XEN KVM


8/7/2012

11

Component CompositionsSC2

M DF

L

SC6

M D F L

SC11

M F D L

Rusle 2 Model Fastest Compositions

• Determining fastest compositions F

Service Isolation Overhead

g p•Not intuitive • Testing / prediction required

• Service Isolation •Was not fastest• Adds overhead

Preliminary Results 41XEN KVM


• Results in maximum hosting costs (# of VMs)

Component CompositionResource Utilization Diversity

Preliminary Results 42Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

Component CompositionResource Utilization Diversity

• Resource utilization varies• Component/VM placements• VM memory size allocations• Hypervisor type KVM/XEN

• Testing required to identify resource utilization• Intuition is insufficient

Preliminary Results 43Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012 Preliminary Results 44Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

8/7/2012

12

August 7, 2012 Wes J. Lloyd PHD Dissertation Proposal Defense Preliminary Results 45 August 7, 2012 Wes J. Lloyd PHD Dissertation Proposal Defense Preliminary Results 46

Scaling Infrastructure requires tuning application parameters in response to available resources.

August 7, 2012 Wes J. Lloyd PHD Dissertation Proposal Defense Preliminary Results 47 48Wes J. Lloyd PHD Dissertation Proposal DefenseAugust 7, 2012

outline - university of washingtonfaculty.washington.edu/wlloyd/slides/prelim.pdf8/7/2012 1 phd...

Documents