Presentation is loading. Please wait.

Presentation is loading. Please wait.

IWQoS 2006 19 -21 June 2006 New Haven, CT Gargi B Dasgupta, Koustuv Dasgupta, Amit Purohit, Balaji Viswanathan, IBM, India Research Lab Grid as a utility-based.

Similar presentations


Presentation on theme: "IWQoS 2006 19 -21 June 2006 New Haven, CT Gargi B Dasgupta, Koustuv Dasgupta, Amit Purohit, Balaji Viswanathan, IBM, India Research Lab Grid as a utility-based."— Presentation transcript:

1 IWQoS June 2006 New Haven, CT Gargi B Dasgupta, Koustuv Dasgupta, Amit Purohit, Balaji Viswanathan, IBM, India Research Lab Grid as a utility-based service Resources offered by grid providers and applications (jobs) executed by consumers QoS-GRAF: Resource Provisioning Framework for Next Generation Grids SLA-based service differentiation Advanced reservations to handle failures QoS-GRAF: A Framework for QoS-based Grid Resource Allocation with Failure Provisioning Better QoS Levels For Medium/Large Jobs 25$/Mbps upto 1 Gbps Incremental Bandwidth 1500$ for 10 MbpsBase Bandwidth 100$/MB upto 1 GB Incremental Storage 1000$ for 200 MBBase Storage RevenueResource Demand Revenue Storage Bandwidth Revenue Demand Revenue 50, , , , , , 1000 Primary Revenue = 1000$ Primary Revenue = 1000$ 200 Units Primary Job 2 Primary Job 1 50 Units Available 50 Units Available R1 R2 R3 100 Units Shared Backup (Job1,Job 2) Objective: Business revenue maximization based on service differentiation Physical Resource Pool Resource Provisioning Manager Batch of Jobs Admission Control Batch Scheduler SLA DB Primary Allocator Backup Allocator SLA Manager Application level objectives Resource level Objectives Monitoring Information Physical Resource Pool Resource Provisioning Manager Admission Control Batch Scheduler SLA DB Primary Allocator Backup Allocator SLA Manager Application level objectives Resource level Objectives Admission Control Admission Control Batch Scheduler Batch Scheduler Utility DB Primary Allocator Backup Allocator SLA Manager Application level objectives Resource level Objectives QoS-GRAF Controller QoS-GRAF Architecture REDUCE OVERBOOKING MINIMIZE REVENUE LOSS Approximat e with discrete step function Worst Case Loss : 250$ Increased $$ SLA based Allocation + Failure Reservation Prototype Implementation SLA-driven Scheduling Integrate with Grid Middleware Backup QoS closer to Primary QoS Level Minimize ∑ k ∑ δ α δ (k) MLBA Maximize ∑ i ∑ j ∑ k ∑ δ U δ (i, j, k)*R δ (i, k) MRPA Backup capacities must not be exceeded Disjoint primary and backup for each resource Jobs that do not have primary allocations on same resource can share backup resources No job is allocated a backup at a QoS level higher than its primary Resource capacities must not be exceeded All jobs must be assigned primary and backup Each job is allocated on a resource at a single QoS level INPUT (1) Batch of K jobs (2) SLA specifications for each resource dependencies (CPU,network,storage,DB) (3) SLA captured by a discrete step function with multiple QoS levels (4) Available resource capacities Job k, Resource j, QoS level i, Dependency δ U δ (i, j, k) : Indicator variable R δ (i, k) : Revenue α δ (k) : Loss (difference of primary and backup revenues) Maximum Revenue Primary Allocation (MRPA) AND Minimum Loss Backup Allocation (MLBA) MLBA SAVES 20% LOSS IN REVENUE MRPA GIVES 30% GAIN IN REVENUE NEAR-OPTIMAL PERFORMANCE SCALABLE WITH LARGE BATCHES SMALL JOBS MEDIUM JOBSLARGE JOBS QoS LEVELS SMALL JOBS MEDIUM JOBS LARGE JOBS QoS LEVELS


Download ppt "IWQoS 2006 19 -21 June 2006 New Haven, CT Gargi B Dasgupta, Koustuv Dasgupta, Amit Purohit, Balaji Viswanathan, IBM, India Research Lab Grid as a utility-based."

Similar presentations


Ads by Google