Transcript datacenters
University of Massachusetts Prof. Prashant Shenoy Upendra Sharma Rahul Singh Gabriel Tarasuk-Levin Timothy Wood Modern Data Centers Research Goals Data centers consist of server clusters with fast network interconnects running diverse applications. A data center may host thousands of applications with varying traffic loads. Improve data center efficiency and simplify management through virtualization. Develop analytic models of application resource consumption and performance. Automate resource management and system monitoring tasks to eliminate error prone, labor intensive administration tasks. Must allocate resources to each application in order to meet Service-level Agreements (SLAs) that describe performance guarantees. Application Workload-mix Provisioning Engine Workload Predictor VM VM VM VM VM VM Monitor and Adjust: Monitor resource utilization of VMs and adjust RAM/CPU resources based on workload requirements. Migrate to avoid hotspots: Automatically detect server overload and transparently migrate VMs between physical servers. Server Consolidation: Virtual machines must be carefully placed on physical servers VM Replication: We are exploring how virtual machine replication can add a new dimension to resource provisioning. Solid State Storage: Solid state flash drives can be used to allow virtual machines to be saved and restored very quickly, allowing for rapid server initialization. … Application Tier VM Tier 2 Workload-mix Classifier VM based Data Center: Each service is run in a VM. Each server in the data center runs multiple VMs. VM Tier 1 Workload-mix System Model Virtualization in Data Centers … VM VM VM … VM Dynamic Provisioning of Multi-tier Applications Model Based Provisioning: Combines workload and system models to guide dynamic resource allocation for multi-tier applications. Workload Classifier: Identifies key attributes of input requests which can be used for clustering and thereafter classification of request classes. System Model: Predicts average response-time for a given workload mix. Workload Predictor: Considers short-term and long-term variations in load and historical-patterns to estimate future workload characteristics. Provisioning Engine: Alters the server capacity according to the forecast made by the capacity planner. Monitoring and Modeling System Monitoring: Data centers host thousands of applications, requiring large scale, low overhead monitoring. Data Mining Application Logs: Application logs can be mined for clues about application activities and workload characteristics. Application Modeling: We create models which relate application resource utilization and performance to incoming workload characteristics. Resource Prediction: The application models can be used to answer “what if” scenarios about how application performance will change under hypothetical workloads. Modeling Virtualized Applications: We have developed models of how virtualization overheads vary based on the types of I/O performed by an application.