Presentation is loading. Please wait.

Presentation is loading. Please wait.

Xavier León PhD defense 5-7-2013.

Similar presentations


Presentation on theme: "Xavier León PhD defense 5-7-2013."— Presentation transcript:

1 Xavier León xleon@ac.upc.edu PhD defense 5-7-2013

2 Outline Shared computational infrastructures Tragedy of the commons PlanetLab as case study Computational markets What is mising? Main contributions Problem Approach Results 2

3 Contributions 3 Model and analysis of a large scale infrastructure Xavier León, Tuan Anh Trinh, and Leandro Navarro. Modeling resource usage in planetary-scale shared infrastructures: PlanetLab's case study. Comput. Netw., Vol. 55 (15), pp. 3394-3407 (2011). ISSN: 1389- 1286. Impact Factor: 1.176 (JCR 2010) External regulation entity to improve load balancing Xavier León, Tuan Anh Trinh, and Leandro Navarro. Using economic regulation to prevent resource congestion in large-scale shared infrastructures. Future Generation Computer Systems, Vol. 26 (4), pp. 599 - 607 (2010). ISSN: 0167-739X. Impact Factor: 2.365 (JCR 2010) Resource provider as regulator to improve energy efficiency Xavier León, and Leandro Navarro. Limits of Energy Saving for the Allocation of Data Center Resources to Networked Applications. In proceedings of the The 30th IEEE International Mini-conference on Computer Communications (IEEE INFOCOM'11), pp. 216-220 (2011), Shanghai, China. ISSN: 0743-166X. Xavier León, and Leandro Navarro. Stackelberg Game to Derive the Limits of Energy Saving for the Allocation of Data Center Resources. Future Generation Computer Systems, Vol. 29 (1), pp. 74-83 (2012). ISSN: 0167-739X. Impact Factor: 2.371 (JCR 2011) User self-regulation for energy efficiency and dynamic allocation Xavier León and Leandro Navarro. Incentives for Dynamic and Energy-aware Capacity Allocation for Multi- tenant Clusters. Submitted to International Conference on the Economics of Grids, Clouds, Systems, and Services. [P1] [Ch 5] [Ch 6] [Ch 7] [Ch 8] [P2] [P3] [P4] [P5]

4 4 Where is the problem?

5 The “Tragedy of the Commons” 5 The benefits of exploitation of a finite resource are perceived by individuals while the costs are shared by the group

6 Shared computational infrastructures 6

7 7 Case study: Planetlab * Data from Codeen Host Status 03-08-2008 Consumption vs Contribution Cumulative Load of nodes

8 Computational markets as a solution 8 Good at matching demand and supply Effectively summarize information through prices User-oriented priorities when demand exceeds supply

9 Gaps on current research 9 Mainly focused on utility optimization User efficiency Profit optimizing Lack of control over externalities Multiple system-wide objectives What about … Load balancing Energy-efficiency

10 Problem statement 10 How to incorporate multiple system-wide objectives into computational markets?

11 11 How do we approach the problem? Contributions

12 [P2] [Ch 6] 12

13 Problem Overexploitation of resources 13

14 Regulative Taxes 14 Price 20 Target Load Tax 16 6 0 00 +4 +2 -2 +2 +0

15 Some results 15 Same efficiency Better load balance Small loss on fairness

16 16 Conclusions Resource hot spots may be present ‣ Need a mechanism to internalize this problem into the market ‣ Influence users’ choice Resource taxes through 3 rd party regulation entity ‣ Suitable in decentralized scenarios with no central control Improves load balancing at the cost of fairness ‣ Same efficiency ‣ Utility uniformity and Envy-freeness reduced

17 [P3] [P4] [Ch 7] 17

18 18 Sharing model Proportional share in datacenters x 1 = 1 x 2 = 2 x 3 = 1

19 Problem Increasing energy costs 19 Power (Watts) Server utilization Idle power usage Active power usage Good for Energy Proportionality  Still high idle power!

20 20 Problem Increasing energy costs Given a specific user demand, how many resources should the resource provider shut down without reducing the QoS?

21 Let users optimize utility… 21 Node state (on/off) Minimum share Minimum parallelism Proportional share model of competition

22 …and provider optimize profit by minimizing energy related costs 22 Node state (on/off) Cost of resource Price paid by users Every user is satisfied Revenue – cost model

23 How to solve it? Stackelberg game 23 Resource Provider Leader / Resource supply Regulator Selfish users Followers How many resources to shutdown? Given k available resources, is every user satisfied? … User demand ? Sorted by power consumption (ascending) Optimal # of resources found! Shutdown remaining resources

24 Some results 24

25 25 Significant costs related to energy consumption ‣ Idle power usage of IT equipment becomes a key factor Resource provider in control of resource supply ‣ Stackelberg model of competition ‣ Suitable when resource provider has full control over supply Upper bound on energy savings ‣ Minimize switched on machines while satisfying users’ requirements Conclusions

26 [P5] [Ch 8] 26

27 Fair share/Capacity scheduler Hadoop 27 Pool of resources (time slots in Hadoop) 1/n Spare resources “Fixed” resource allocation

28 Dynamic shares market 28 Pool of resources (time slots in Hadoop) Capacity saving Spare resources that can Be shutdown Reused by others

29 29 Dynamic shares market Application A Application B Application C Cluster Savings proportional to Time and share not used No credits -> default to fair share Allow more than fair share if enough credits If competition -> Proportional share

30 Some results 30

31 31 Issues we address ‣ No control over share guarantees ‣ Lack of dynamic and elastic resource allocation ‣ No notion of energy-related costs Provide incentives for user self-regulation ‣ Decentralize decision making Dynamic and energy-aware resource allocation ‣ Incentive compatible Conclusions

32 32 Wrapping up Conclusions and Future work 32

33 33 Conclusions How to incorporate multiple system-wide objectives into computational markets?

34 34 Future work What about other system-wide metrics? ‣ Reward the social benefit of executions ‣ Control monopolies on federated infrastructures Explore these mechanisms on real deployments ‣ Are they feasible? ‣ Easy of use? ‣ Need for computer-aided decision making?

35 Xavier León xleon@ac.upc.edu PhD defense 5-7-2013


Download ppt "Xavier León PhD defense 5-7-2013."

Similar presentations


Ads by Google