Dynamic Resource Allocation for Shared Data Centers using Online Measurements Abhishek Chandra, Weibo Gong, Prashant Shenoy UMASS Amherst http: //lass. cs. umass. edu Data Centers Goals: §Server farms §Rent computing and storage resources to applications §Revenue for meeting Qo. S guarantees §Satisfy application Qo. S guarantees §Maximize resource utilization of platform §Robustness against “Slashdot” effects Dynamic Resource Allocation §Estimate workloads for near future §Periodically re-allocate resources among applications Measured Usage PREDICTOR Expected Load APPLICATION MODELS MONITOR Challenges: ALLOCATOR System Metrics §Reallocation at short time-scales §No prior workload profiling/knowledge §Low overhead Rsrc Reqmts Resource Shares RESOURCE Prediction Application Model Allocation Time Series Analysis Measurement-based Model Utility Model Workload as time series Discontent function for Qo. S goal violations Predicted Workload Parameters History Adaptation Window Qo. S Reqmts Model Rsrc Reqmts Discontent Di Goal Response Time Prediction Application Model Allocation AR(1) Model Transient Queuing Model Non-linear optimization subject to aj : number of arrivals in jth interval R : Autocorrelation function ej : White noise Relation between mean response time and share Simulation Results: Using Soccer World Cup’ 98 Web Traces Workload Prediction Time (min) Share Allocation System Discontent