A Case for Economy Grid Architecture for Service

A Case for Economy Grid Architecture for Service Oriented Grid Computing Authors: Rajkumar Buyya, David Abramson & Jonathan Giddy Presenter: Diego Lopez Agnostic: Djuradj Babic June 12, 2006

Outline § Introduction § Grid Economy and Resource Management Issues § Economy Models and Related Work § GRACE § Resource Trading and Scheduling Experimentation § Conclusion and Future Work June 12, 2006 2

1. Introduction “ We expect that an economy driven approach to resource management and scheduling will make a great impact on the eventual success and widespread adoption of the Grid in day-to-day computational activities. ” Dr. Buyya June 12, 2006 Researcher Giddy Prof. Abramson Source: Buyya, R http: //www. buyya. com 3

1. Introduction § Grid environment is complex ($$$) § Different access cost models § Dynamically varying loads and availability conditions § Use of economic models in the Grid to encourage participation and wide-scale adoption § Proposal of computational economy framework that leverage existing Grid sites June 12, 2006 4

View of Economic Grid June 12, 2006 Source: Buyya, R A Case for Economy Grid Architecture for Service Oriented Grid Computing (Pg. 2) 5

2. Grid Economy and Resource Management Issues § Establish policies that promote Grid resource sharing § 2 key players in Grid economy § Resource providers (GSP) § Resource consumers (GRB) § Consumers interact with brokers to express their budget and deadline requirements from the Grid June 12, 2006 6

Proposal of GRACE § Grid Architecture for Computational Economy § Leverage of existing infrastructures: § Globus/Legion § Condor/G § Provide an infrastructure that allows for: § Info/Market directory for publicizing entities § Model for determining value of resources § Resource pricing schemes § Accounting, Billing and Payment mechanisms June 12, 2006 7

3. Economy Models and Related Work Possible economic models for resource trading and pricing strategies § Commodity Market § Posted Price § Bargaining § Tendering/Contract-Net § Auction § Bid-based Proportional Resource Sharing § Community/Coalition/Bartering June 12, 2006 8

Examples of Computational Economy Systems SYS_NAME Mariposa (’ 96) Mungi (’ 02) Popcorn (’ 98) ECO_MODEL Bidding Commodity Auction PLATFORM Dist. Dbase Storage servers Web browser Mojo Nation Credit-based or bartering Qo. S based computational Network storage Web browser (applets) Java Market June 12, 2006 9

4. GRACE § Use of well-adopted Grid technologies, Globus/Condor § Development of middleware services for resource trading using different economic models § Development of advanced user-centric Grid resource brokers June 12, 2006 10

4. 1 Grid Resource Broker (GRB) § Mediator between user and grid resources § *Nimrod – parametric modeling language § Use of Nimrod/G broker (superscheduler) § Job Control Agent § Schedule Advisor § Grid Explorer § Trade Manager § Deployment Agent June 12, 2006 http: //ipdps. cc. gatech. edu/2000/papers/Abramso n. pdf 11

RM: Local Resource Manager, TS: Trade Server Nimrod/G Broker Nimrod/G Client Nimrod/G Engine Schedule Advisor Trading Manager Grid Store Grid Dispatcher Grid Explorer Grid Middleware Globus, Legion, Condor, etc. TM TS GE GIS Grid Information Server(s) RM & TS G L G Globus enabled node. June 12, 2006 RM & TS L Legion enabled node. Src: http: //www. buyya. com/ecogrid/ C Condor enabled node. 12

4. 2 Economy Grid Middleware in Globus Context § Trade Server (TS) – maximize the resource utility and profit for its owner § Pricing Policies – define prices for resources based on economic models previously mentioned § Resource Accounting and charging – tracking resource usage for billing and auditing purposes June 12, 2006 13

4. 3 Grid Open Trading Protocols and Deal Template § Establish rules and format for exchanging commands between a GRACE client (Trade Manager) and Trade Server § Deal Template (DT) contains § CPU time § Storage requirements § Initial offer § This trading overhead can be reduced if prices are announced via GIS June 12, 2006 14

4. 4 Pricing, Accounting, and Payment Mechanisms § N-ways to determine resource pricing § Fixed price model (no Qo. S like today’s www) § Usage timing (peak, off-peak) § Bulk purchase § Demand supply § Loyalty of customers (i. e. frequent flyer miles) § Calendar based June 12, 2006 15

4. 4 Pricing, Accounting, and Payment Mechanisms § Service items to be charged § CPU time § Memory § Storage used § Software and Libraries accessed (ASP) § Access to these services can be charged § Individually § Combination (costing matrix) June 12, 2006 16

4. 4 Pricing, Accounting, and Payment Mechanisms § Prepaid – purchase credits from GSP or Grid Bank § Use and pay later (like electricity) § Pay as you go (wireless calling cards) § Grants based § *Billing services handled by 3 rd party: § Net. Cheque § Paypal June 12, 2006 *not incorporated into GRACE described in this paper 17

4. 5 System Prototype & Demo Experiences § Prototype of the Nimrod/G resource brokering demo held during HPDC 2000 § Parameter study experiment performed over Grid resources located in both Australia and the US § Ability to change deadline and budget to trade-off cost vs. timeframe to illustrate Grid marketplace dynamics June 12, 2006 18

5. Resource Trading and Scheduling Experimentation § Experiment to test operation of Grid Trade Server across 5 systems (165 jobs) § Use of Posted Price Market Model for the Nimrod/G brokering § Runs during peak time vs. off-peak time § Access price expressed in Grid units per CPU second (G$) § Resource/service provided by GRACE framework June 12, 2006 19

Economy Grid Results § Cost-Optimization algorithm successfully § Minimized artificial access cost per resource § Completed within one-hour deadline § Initial calibration phase ensures completion within budget/time constraints § Scheduler excluded usage of resources during peak time § Scheduler predictions met deadline using least-expensive resources available June 12, 2006 20

6. Conclusion and Future Work § GRACE leverages existing middleware systems (Condor/Legion/Globus) § Nimrod/G can discover best resource providers based on user’s requirements § Nimrod/G does not support dynamic prices once initial scheduling is proposed § Nimrod/G Portal available … June 12, 2006 21