Dynamic Resource Allocation for Shared Data Centers Using

Dynamic Resource Allocation for Shared Data Centers Using Online Measurements Abhishek Chandra Weibo Gong Prashant Shenoy UMASS Amherst http: //lass. cs. umass. edu/projects/shop UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science

Motivation n Data Centers n n Server farms Rent computing and storage resources to applications Revenue for meeting Qo. S guarantees Goals: n n n Satisfy application Qo. S guarantees Maximize resource utilization of platform Robustness against “Slashdot” effects UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 2

Dynamic Resource Allocation n Periodically re-allocate resources among applications Estimate resource requirements for near future n Challenges: n n n Reallocation at short time-scales No prior workload profiling/knowledge Low overhead Approach: Online Measurement-based Allocation UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 3

Talk Outline ü Motivation n System Model n Dynamic Allocation Techniques n Experimental Results n Conclusions UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 4

Resource Model Resource GPS n n n Queuing System Generalized Processor Sharing (GPS) scheduler Request classes n n Different arrival processes, service time distributions Qo. S Goal: Mean Response Time UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 5

Dynamic Resource Allocation Measured Usage PREDICTOR Expected Load MONITOR APPLICATION MODELS Rsrc Reqmts ALLOCATOR System Metrics Resource Shares RESOURCE UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 6

Dynamic Resource Allocation Measured Usage PREDICTOR Expected Load MONITOR APPLICATION MODELS ALLOCATOR System Metrics RESOURCE UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 7

Monitoring n Measure system and application metrics n n n Queue lengths Request response times Monitoring windows Measurement Interval History Adaptation Window Time UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 8

Prediction n Short-term prediction of workload characteristics n n n Request arrival rate Average service time Use history of measured system metrics Mean AR(1) Last value History Adaptation Window UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 9

Prediction Accuracy Workload Prediction Error Time (min) UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 10

Dynamic Resource Allocation PREDICTOR Expected Load MONITOR APPLICATION MODELS Rsrc Reqmts ALLOCATOR Resource Shares RESOURCE UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 11

Measurement-based Model n n n Goal: Relate Qo. S metric to resource requirement Idea: Model parameterized by online measurements Advantages: n n n Parameters do not need to be computed Allow adaptation to dynamic workload Proposed: Transient Queuing System Description UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 12

Transient Queuing Model n n Transient queuing behavior over adaptation window Relation between mean response time T¯ and application share w n n Little’s Law: Relation is parameterized by the measured workload n Arrival rate λ and mean service time s¯ UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 13

Resource Allocation: Utility Model u 1 Optimization u 2 n n Discontent function: Measures the Qo. S violations of an application Constrained Optimization problem UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 14

Constrained Optimization Formulation Discontent Di Goal Response Time n Non-linear Optimization Problem: subject to n Solved using Lagrange multiplier method UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 15

Talk Outline ü Motivation ü System Model ü Dynamic Allocation Techniques n Experimental Results n Conclusions UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 16

Experimental Setup n Simulation experiments n Soccer World Cup’ 98 Traces n Results based on a 24 -hour portion of the trace n n n 755, 000 requests Mean req rate: 8. 7 req/sec Mean req size: 8. 47 KB UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 17

Adaptation to Transient Overloads Workloads Share Allocation Shares adapt to changing workload characteristics UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 18

Adaptation: System Discontent GPS without reallocation GPS with reallocation System Discontent is lowered substantially UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 19

Conclusions n n Dynamic Resource Allocation needed for data centers Measurement-based allocation: n n n Monitoring and Prediction gather online state Use this state for application modeling and allocation Future Work: n n Prediction policies Utility functions http: //lass. cs. umass. edu/projects/shop UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 20

Related Work n Prediction n n Application Models n n n Statistical Prediction Models [Zhang 00] Queuing-Theoretic Models [Carlstrom 02, Liu 01] Control-Theoretic Models [Abdelzaher 02, Lu 01] Data Centers n n n MUSE [Chase 01] COD [Moore 02] Oceano [Appleby 01] UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 21