Resource Provisioning based on Lease Preemption in InterGrid Mohsen Amini Salehi, Bahman Javadi, Rajkumar Buyya Cloud Computing and Distributed Systems.

Resource Provisioning based on Lease Preemption in InterGrid

Mohsen Amini Salehi , Bahman Javadi, Rajkumar BuyyaCloud Computing and Distributed Systems (CLOUDS) Laboratory,

Department of Computer Science and Software Engineering,The University of Melbourne, Australia

Mohsena,bahmanj,[email protected]

• Providing computational resources for users is one of the challenges in the high performance computing.

• Resource Providers (RP)?

• Grid 5000, DAS-2,Amazon EC2,etc.

Introduction

InterGrid

• provides an architecture and policies for inter-connecting different Grids.

• Computational resources in each RP are shared between grid users and local users.

• Provisioning rights of the resources in a Grid are delegated to the InterGrid Gateway (IGG).

• Local users vs Grid (External) users.

Lease based Resource Provisioning in InterGrid

• A lease is an agreement between resource provider and resource consumer whereby the provider agrees to allocate resources to the consumer according to the lease terms presented.

• Virtual Machine (VM) technology is a way to implement lease-based resource provisioning.

• VMs are able to get suspended, resumed, stopped, or even migrated.

• InterGrid makes one lease for each user request.

InterGrid

Problem Statement

• How to provision resources for local requests when existing resources have been allocated

to grid requests?

• Partitioning• Preempting.

Challenges of Preempting• Is that really useful?!

• Originally, it is not allowed to preempt leases without permission.– How to do that?– What to do with preempted leases?

• lease preemption has some side-effects:– imposes time overhead– can potentially affect other reservations

Challenges of Preempting…

• In an RP, usually several leases have to be preempted to make sufficient resources. – there are also several choices for preemption!

(Candidate Sets).– candidate sets have various amount of imposed

overhead. Different number of grid users get affected.

• How to choose the optimal candidate set for preemption?

Which one is optimal Candidate set?

42

1

2 62

42

1

2

Related Work

• Haizea: a lease scheduler for advanced reservation and best effort leases. For preemting it just considers the preemptability of the lease.

• Sotomayor et al. estimated the overhead time imposed by preempting a lease (suspending and resuming a VM)

• Walters et al. used preemption to give precedence to interactive jobs inside a cluster. But they focus on how to checkpoint the preempted job, and how to resume the preempted job.

• Kettimuthu et al. applied preemption policy to decrease waiting time.

Proposed Solution(1): make the preemption possible

• We introduce different request types (lease type) in InterGrid. – At the moment, a user request in InterGrid is

composed of: • Virtual Machine (VM) name needed by the user.• Number of VMs needed.• Ready time: the time that requested VMs should be ready.• Wall time: duration of the lease.• Deadline: the time that serving the request must be finished.

– Based on the lease types, it is determined how to schedule the lease and what to do with a preempted lease.

Proposed Solution(1): Introducing Different Lease Types

• Best Effort-Cancelable: – neither guarantee the deadline nor the wall time. – impose the minimum overhead time in preemption.

• Best Effort-Suspendable:– guarantees the wall time but not in a specific deadline. – overhead is the time to suspend a VM, reschedule , and

resume it.

• Deadline Constraint-Migratable:– guarantee both the wall time and deadline of the lease.

• Deadline Constraint-Non-Preemptable:– guarantees both deadline and wall time .

Proposed Solution(2): Preemption Policy-1

• Minimum Overhead Policy (MOV)– aims at maximizing resource utilization. – tries to minimize the time overhead imposed to

the underlying system– preempts a candidate set that leads to the

minimum overhead. – It works out the total overhead imposed to the

system by each candidate set and the set with minimum overhead is selected.

Proposed Solution(2): Preemption Policy-2

• Minimum Leases Involved Policy(MLIP)– Users do not like that their leases get affected by

preemption.– As a user centric policy, MLIP tries to satisfy more

users by preempting less number of leases.– In this policy a candidate set that contains

minimum number of leases is selected from all the candidate sets.

– MLIP disregards the type of leases involved in a candidate set.

Proposed Solution(2): Preemption Policy-3

• Minimum Overhead Minimum Lease Policy (MOML)– MOML is a balance between MOV

Minimum Overhead Minimum Lease Policy (MOML)

Performance Evaluation-Metrics

• Local and Grid Request Rejection Rate• Resource Utilization• Number of Lease Preemption

Experiment configuration:

• We use Lublin99 workload model.• We experiment an RP with 32 nodes.

Experimental Results:Local and Grid Request Rejection Rate

Resource Utilization

Number of Lease Preemptions

Conclusion

• we leveraged preempting grid leases in favour of local requests.

• We proposed different typesof leases for lease based resource providers.

• We proposed three policies for lease preemption:– MOV as a policy that improves system utilization,– MLIP that results in less number of preemption and

increasing user satisfaction, – MOML which makes a trade-off between resource

utilization and user satisfaction.

Future Work

• Scheduling policies in IGG that makes less preemption.

• we are interested in optimal sequence of grid leases in a site.

THANK YOUAny Question?

References• Chase, J. S., Irwin, D. E., Grit, L. E., Moore, J. D. &Sprenkle, S. E. (2003),

Dynamic virtual clusters in a grid site manager, in `Proceedings of the 12th IEEE International Symposium on High Performance Distributed Computing', Washington, DC,USA, pp. 90-98.

• De Assunc~ao, M., Buyya, R. & Venugopal, S. (2008), `InterGrid: A case for internetworking islands of Grids', Concurrency and Computation: Practice and Experience 20(8), 997-1024.

• Lublin, U. & Feitelson, D. G. (2001), `The workload on parallel supercomputers: Modeling the characteristics of rigid jobs', Journal of Parallel and Distributed Computing 63, 2003.

• Sotomayor, B., Keahey, K. & Foster, I. (2008), Combining batch execution and leasing using virtual machines, in `Proceedings of the 17th International Symposium on High Performance Distributed Computing', ACM, New York, NY, USA,pp. 87-96.