f19dea4f11f954af36e56b878be58fbc.ppt
- Количество слайдов: 18
Quantifying the Benefits of Resource Multiplexing in On-Demand Data Centers Abhishek Chandra Prashant Shenoy UMASS Amherst Pawan Goyal IBM Almaden, San Jose UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science
Motivation n On-demand Data Centers n n Server farms Rent computing and storage resources to applications Revenue for meeting application workload levels Goals: n n n Satisfy dynamically changing application requirements Maximize resource utilization of the platform Robustness against “Slashdot” effects UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 2
Dynamic Resource Allocation n Existing techniques: n n n Common features: n n Oceano [Appleby 01], HP Utility Data Center [Rolia 00], MUSE [Chase 01], COD [Doyle 02], SHARC [Uragaon 02] Differ in allocation policies and mechanisms Periodically re-allocate resources among applications Estimate workloads for near future Statistical multiplexing of resources Question: Which techniques work best and when? UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 3
On-demand Allocation: Practical Issues n n How often and how fine should the re-allocation be done? How well can the application requirements be estimated? How much “head room” should be allowed to absorb transient loads? Do large number of customers lead to better statistical multiplexing? UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 4
Talk Outline ü Motivation n System Model and Metrics n Performance Study n Conclusions and Future Work UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 5
System Model n n Cluster of servers Homogeneous pool of resources No constraints on application placement Time granularity (Δt): Period of re-allocation n n E. g. : re-allocate once every minute, hour, day Space granularity (Δs): Resource allocation unit n E. g: re-allocate partial/whole server, server group UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 6
Optimal Resource Allocation n Infinitesimally small allocation granularity Allocates precise amount of resource No resource wastage Ropt Resource Allocation Time UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 7
Practical Resource Allocation n Allocation done periodically and in fixed quanta Fixed resource allocation for next period Clairvoyant scheme: Predict peak application requirements for the next allocation period Δs Δt Resource Allocation Time UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 8
Capacity Overhead Rpract ρ Ropt Resource Allocation Time UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 9
Performance Study n Workload: n 3 e-commerce traces n 24 -hour long Workload Number of Requests Avg. Request Size Peak bit-rate Ecommerce 1 1, 194, 137 3. 95 KB 458. 1 KB/s Ecommerce 2 1, 674, 672 3. 85 KB 1631. 0 KB/s Ecommerce 3 251, 352 7. 24 KB 1346. 9 KB/s UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 10
Effect of Allocation Granularity Time granularity Space granularity § Fine time scale with reasonably fine resource unit desirable UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 11
Effect of Prediction Inaccuracy n Fine allocation is better even with inaccurate prediction UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 12
Effect of Overprovisioning n Finer allocation achieves same “head room” with less overhead UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 13
Effect of Number of Customers n Large number of customers provide more opportunity for statistical multiplexing UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 14
Data Center Architectures n Dedicated n n n Shared n n n Allocation of whole servers Typical reallocation in order of 30 minutes Fractional server resources Reallocation in seconds or minutes Fast Reallocation n n Reserved server pools, remote booting Reallocation in a few minutes UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 15
Comparison of Architectures Data Center Configuration Number of customers Optimal Reqmt (Num of servers) Dedicated Architecture (Num of servers) Fast Reallocation (Num of servers) Shared Architecture (Num of servers) Small 3 20 34 31 25 Medium 15 100 388 304 148 Large 30 1000 5017 3759 1739 UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 16
Implications and Opportunities n Cost of re-allocation n n Partial server: ~1 syscall/min Full server: Rebooting, disk scrubbing, etc. Virtual machines: Low cost of reallocation with encapsulation Prediction: n n Work-conserving scheduler at fine time-scales Accurate prediction possible at minutes, hours UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 17
Conclusions and Future Work n n Dynamic Resource Allocation for data centers Fine allocation granularity desirable n n Even with inaccurate prediction To achieve more “head room” Large number of customers lead to higher multiplexing benefits Future Work: n n n Effect of affinity, placement constraints Re-allocation overhead Stability of resource allocation UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science 18
f19dea4f11f954af36e56b878be58fbc.ppt