Outline Introduction Challenges Background Research Questions - PDF document

10/22/2016 Outline � Introduction � Challenges � Background � Research Questions � Methodology � Research Results � Performance Modeling for Component Composition Wes J. Lloyd � Noisy Neighbor Detection October 22, 2016 � Workload Cost Prediction Methodology � Summary � Future Directions Institute of Technology, University of Washington- Tacoma 2 Cloud Computing Cloud Computing NIST General Definition Stack “Cloud computing is a model for enabling convenient, Software on-demand network access to a shared pool of configurable computing resources (networks, servers, storage, applications and services) that can be rapidly Platform provisioned and reused with minimal management effort or service provider interaction”… Infrastructure 3 4 Cloud Computing Cloud Computing Stack Stack User manages: User manages: User manages: Application Services Application Services, Application Services, PaaS Application Infrastructure, Application Infrastructure, Virtual Servers Virtual Servers IaaS IaaS 5 6 1

10/22/2016 Cloud Computing Cloud Computing Stack Stack SaaS SaaS User manages: User manages: User manages: User manages: Application Services Application Services Application Services, Application Services, Application Infrastructure, PaaS Application Infrastructure, PaaS IaaS Virtual Servers Virtual Servers IaaS IaaS 7 8 Microprocessors Advancements Virtualization � Smaller die sizes (microns) � Lower voltages � Improved heat dissipation � Energy conservation � More transistors, but with similar clock rates � How do we harness this new transistor density? � Multicore CPUs � Improve computational throughput � How do we utilize many-core processors? 9 10 Virtualization Containerization Virtualization Containerization 11 12 2

10/22/2016 Public Cloud Example: Traditional Application Deployment Netflix App � Amazon Elastic Compute Cloud (EC2) Data Business Logging Server � Continuously run 20,000 to 90,000 VM instances � Across 3 regions Spatial DB � Host 100s of microservices Services Apache rDBMS redis � Process over 100,000 requests/second Tomcat Services � Host over 1 billion hours of monthly content Physical DODB / Server(s) Services NOSQL 13 Outline � Introduction � Challenges � Background � Research Challenges � Methodology � Research Results � Performance Modeling for Component Composition � Noisy Neighbor Detection � Workload Cost Prediction Methodology � Summary � Future Directions 15 16 VM-image VM-image Research Challenges – WHERE VM Physical Host tomcat postgresql VM VM nginx VM-image VM-image postgis Service Isolation Where should we Physical Host haproxy redis-server VM provision? Application “Stack” VM-image VM-image Physical Host VM VM Physical Host 17 18 3

10/22/2016 VM-image VM-image VM-image VM-image tomcat postgresql tomcat postgresql VM haproxy nginx VM postgis Physical Host Physical Host VM VM VM VM VM-image VM-image VM-image VM-image Component nginx haproxy Physical Host Physical Host Composition VM VM VM-image VM-image Physical Host VM-image VM-image Physical Host postgis redis-server redis-server VM VM VM VM Physical Host Physical Host 19 20 VM-image tomcat haproxy VM Physical Host Physical Host VM VM VM-image Provisioning postgresql Elasticity Physical Host Physical Host nginx Variation postgis VM Physical Host Physical Host VM-image redis-server VM VM Physical Host Physical Host 21 22 Physical Host Physical Host Server Multi-tenancy Physical Host Physical Host Consolidation Physical Host Physical Host Physical Host Physical Host 23 24 4

10/22/2016 Research Challenges – WHERE Service Isolation Component Composition Physical Host Provisioning Server Consolidation Variation Overprovisioning Physical Host Multi-tenancy Overprovisioning Physical Host Resource Contention Physical Host 25 26 Research Challenges - WHAT Research Challenges - WHAT Size Quantity Size Quantity Vertical Scaling Horizontal Scaling Vertical Scaling Horizontal Scaling Amazon Qualitative VM types Resource descriptions VM VM What should we m1.large VM VM VM VM VM VM VM VM Virtualization c3.xlarge Virtualization m2.2xlarge VM VM VM VM VM VM VM VM Overhead Hypervisors provision? m4.2xlarge c1.xlarge VM VM VM VM VM VM VM VM m3.medium m2.4xlarge VM VM VM VM VM VM VM VM d2.8xlarge d2.8xlarge d2.8xlarge d2.8xlarge m1.xlarge c3.large Performance Performance 53+ types 27 28 Research Challenges - WHAT Research Challenges - WHAT Size Quantity Size Quantity Vertical Scaling Horizontal Scaling Vertical Scaling Horizontal Scaling Qualitative Amazon Qualitative Amazon VM types VM types Resource descriptions Resource descriptions m1.large c3.xlarge Virtualization Virtualization Virtualization m2.2xlarge Hypervisors Overhead Hypervisors m4.2xlarge c1.xlarge m3.medium m2.4xlarge d2.8xlarge d2.8xlarge d2.8xlarge d2.8xlarge m1.xlarge Performance c3.large 53+ types 29 30 5

10/22/2016 Research Challenges - WHEN Research Challenges - WHAT Size Quantity Vertical Scaling Horizontal Scaling Amazon Qualitative VM types When should we Resource descriptions Virtualization Virtualization provision? Overhead Hypervisors Performance 32 31 Outline Research Challenges - WHEN � Introduction � Challenges Hot Spot Detection Provisioning Latency � Background � Research Questions Future Load Prediction Pre-provisioning � Methodology � Research Results � Performance Modeling for Component Composition � Noisy Neighbor Detection � Workload Cost Prediction Methodology � Summary � Future Directions 33 33 34 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: � Disk read/write throughput � Wide variance of implementations � Network read/write throughput � Systems continuously evolve � PM capacities vary dynamically NP-Hard � Performance modeling (large problem space) � VM resource utilization varies � Virtualization misunderstood or overlooked � Component requirements vary 35 36 6

10/22/2016 Outline Research Questions (1/2) � Introduction RQ-1: Component composition � Challenges � Background How does resource utilization and service oriented � Research Questions application (SOA) performance vary relative to � Methodology component composition across VMs? � Research Results � Performance Modeling for Component Composition RQ-2: Performance modeling � Noisy Neighbor Detection Which resource utilization variables and modeling � Workload Cost Prediction Methodology techniques best help predict SOA performance? � Summary � Future Directions 37 38 Outline Research Questions (2/2) � Introduction RQ-3: Noisy neighbors � Challenges � Background What performance implications result from � Research Questions resource contention and how can we avoid it? � Methodology � Research Results RQ-4: Infrastructure prediction � Performance Modeling for Component Composition � Noisy Neighbor Detection How can we predict the required cloud � Workload Cost Prediction Methodology infrastructure to satisfy performance � Summary requirements for SOA workload hosting? � Future Directions 39 40 Scientific Modeling Workloads Methodology � USDA Cloud Services Integration Platform (CSIP): � Benchmark Workloads � Framework for scientific modeling-as-a-service � Scientific Modeling Workloads � Scientific modeling SOAs: � Profile resource utilization � RUSLE2 – Soil erosion model � Collect VM-level data � WEPS – Wind Erosion Prediction System � Analytics: construct performance and cost models � SWAT-DEG: Stream channel degradation prediction � R: statistical regression, neural networks Monte carlo workloads � Comprehensive Flow Analysis tools � Evaluate and refine models Load estimator, Load duration curve, Flow duration � Develop heuristics Curve, Baseflow, Flood analysis, Drought analysis 41 42 7