10/26/2014 Outline � Introduction (VM) Virtual Machine � Research goals (PM) Physical Machine � Challenges � Research questions � Background � Research contributions Ph.D. Dissertation Defense � Supporting Infrastructure � Research Results � Performance Modeling for Component Composition Wes J. Lloyd � VM Placement to Reduce Resource Contention October 27, 2014 � Workload Cost Prediction Methodology � Conclusions Colorado State University, Fort Collins, Colorado USA October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 2 Outline Research Goals � Introduction � Research goals � Support application migration: � Challenges VM component composition, dynamic scaling, infrastructure alternatives � Research questions � Background � Maximize: application throughput � Research contributions Requests per second � Supporting Infrastructure � Minimize: hosting costs, server occupancy � Research Results Number of VMs, CPU cores, memory, disk space, hosting costs � Performance Modeling for Component Composition � Minimize response time � VM Placement to Reduce Resource Contention � Workload Cost Prediction Methodology Average service execution time (sec/min) � Conclusions October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 3 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 4 1
10/26/2014 Research Challenges – WHERE Research Challenges – WHERE Service Isolation Component Composition Provisioning Where should Server Consolidation Variation infrastructure be Multi-tenancy Overprovisioning provisioned? Resource Contention October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 5 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 6 Research Challenges - WHAT Research Challenges - WHAT Size Quantity Size Quantity Vertical Scaling Horizontal Scaling Vertical Scaling Horizontal Scaling Amazon Qualitative Amazon Qualitative VM types VM types Resource descriptions Resource descriptions VM VM What infrastructure m1.large m1.large VM VM VM VM VM VM VM VM c3.xlarge c3.xlarge Virtualization Virtualization Virtualization Virtualization m2.2xlarge VM VM VM VM m2.2xlarge VM VM VM VM Overhead Hypervisors Overhead Hypervisors should be provisioned? m1.small m1.small c1.xlarge c1.xlarge VM VM VM VM VM VM VM VM m3.medium m3.medium m2.4xlarge VM VM VM VM m2.4xlarge VM VM VM VM c1.medium c1.medium c1.medium c1.medium c1.medium c1.medium c1.medium c1.medium m1.xlarge m1.xlarge Performance Performance c3.large c3.large October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 7 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 8 2
10/26/2014 Research Challenges - WHEN Research Challenges - WHEN Hot Spot Detection VM Launch Latency When should Future Load Prediction Pre-provisioning infrastructure be provisioned? October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 9 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 10 10 Outline Research Questions (1/3) � Introduction DRQ-2: Performance modeling � Research goals � Challenges What are the most important resource � Research questions utilization variables and modeling techniques � Background for predicting service oriented application (SOA) � Research contributions performance? � Supporting Infrastructure DRQ-3: Component composition � Research Results � Performance Modeling for Component Composition How does resource utilization and SOA � VM Placement to Reduce Resource Contention performance vary relative to component � Workload Cost Prediction Methodology composition across VMs? � Conclusions October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 11 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 12 3
10/26/2014 Research Questions (2/3) Research Questions (3/3) DRQ-4: VM placement implications DRQ-6: Infrastructure prediction When dynamically scaling cloud infrastructure How effectively can we predict required to address demand spikes how does VM infrastructure for SOA workload hosting by placement impact SOA performance? harnessing resource utilization models and Linux time accounting principles? DRQ-5: Noisy neighbors How can noisy neighbors , multi-tenant VMs that cause resource contention be detected? What performance implications result when ignoring them? October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 13 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 14 Virtual Machine (VM) Placement Outline as “Bin Packing Problem” � Introduction � Components items � virtual machines (VMs) bins � Research goals n k � Challenges � Virtual machines (VMs) items � physical machines (PMs) bins 4 15 � Research questions 5 52 � Dimensions � Background 6 203 Bell’s Number � # CPU cores, CPU clock speed, architecture � Research contributions 7 877 � RAM, hard disk size, # cores � Supporting Infrastructure 8 4,140 � Disk read/write throughput � Research Results 9 21,147 � Network read/write throughput � Performance Modeling for Component Composition � PM capacities vary dynamically NP-Hard n . . . � VM Placement to Reduce Resource Contention � VM resource utilization varies � Workload Cost Prediction Methodology � Component requirements vary � Conclusions October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 15 October 27, 2014 Wes J. Lloyd PHD Dissertation Defense 16 4
10/26/2014 Virtual Machine (VM) Placement Why Gaps Exist as “Bin Packing Problem” � Components items � virtual machines (VMs) bins � Public clouds � Virtual machines (VMs) items � physical machines (PMs) bins � Research is time/cost prohibitive � Dimensions � Hardware abstraction: Users are not in control � # CPU cores, CPU clock speed, architecture � Rapidly changing system implementations � RAM, hard disk size, # cores � Private clouds: systems still evolving � Disk read/write throughput � Network read/write throughput � Performance models (large problem space) � PM capacities vary dynamically NP-Hard � Virtualization misunderstood or overlooked � VM resource utilization varies � Component requirements vary October 27, 2014 Wes J. Lloyd PHD Dissertation Defense 17 October 27, 2014 Wes J. Lloyd PHD Dissertation Defense Approaches & Gaps 18 Outline Primary Research Contributions � Introduction � In the context of SOA migration to IaaS Clouds � Research goals � Challenges � Resource utilization modeling to predict � Research questions component composition performance � Background � Research contributions � VM placement improvement to reduce contention � Supporting Infrastructure � Private IaaS: LeastBusy VM placement � Research Contributions � Public/Private IaaS: Noisy-Neighbor Detection, Avoid � Performance Modeling for Component Composition heterogeneous VM type implementations � VM Placement to Reduce Resource Contention � Workload cost prediction methodology for � Workload Cost Prediction Methodology infrastructure alternatives to reduce hosting costs � Conclusions October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 19 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 20 5
10/26/2014 Outline Scientific Modeling Workloads � Introduction � CSIP: USDA platform for model services � Research goals � Challenges � Service oriented application surrogates � Research questions � RUSLE2 – Soil erosion model � Background � Research contributions � WEPS – Wind Erosion Prediction System � Supporting Infrastructure � SWAT-DEG: Stream channel degradation prediction � Research Results Monte carlo workloads � Performance Modeling for Component Composition � Comprehensive Flow Analysis tools � VM Placement to Reduce Resource Contention Load estimator, Load duration curve, Flow duration � Workload Cost Prediction Methodology Curve, Baseflow, Flood analysis, Drought analysis � Conclusions October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 21 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense Research Questions & Methodology 22 VM-Scaler VM-Scaler • REST/JSON Web services application • Harnesses EC2/Eucalyptus API • Provides cloud infrastructure management • Supports scientific modeling-as-a-service • Supports research and IaaS experimentation • Supports Amazon, Eucalyptus 3.x clouds • Extensible to others, e.g. OpenStack future future October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 23 23 October 27, 2014 Wes J. Lloyd PhD Dissertation Defense 24 24 6
Recommend
More recommend
