Cost-Efficient Resource Management for Scientific Workflows on the - - PowerPoint PPT Presentation

Ilia Pietri School of Computer Science, University of Manchester, U.K

Cost-Efficient Resource Management for Scientific Workflows on the Cloud

Scientific Workflows Virtual Machines (VMs) Hosts

User Cloud Provider

Meet User Constraints Goals Provide Energy Efficiency

Overview

1/21

• Scientific Workflows

– DAGs* – Data dependency constraints – Gaps in the schedule

• Goal:

– Increase resource utilisation and achieve cost-efficient provisioning

• Estimation of resource needs
• CPU frequency selection

*DAG: Directed Acyclic Graph

C D E F A G B

Problem Description

2/21

• A wide range of possible configurations

– Number of resources – CPU frequency

• New pricing schemes

– Pricing for CPU provisioning – Faster resources cost more

• Which option to choose?

– Extra resources at a lower frequency – Less resources but faster

3/21

Motivation

4/21

Motivation

Motivation

5/21

• Assuming the user is interested in executing

a scientific workflow

– How many resources to provision? – Cost vs performance User: cost-efficient execution, as quickly as possible Provider: minimize energy costs – Performance modelling estimate workflow execution time (and related costs) for a different number of slots

Determining the Amount of Resources

6/21

• Level-based Estimation Model

– Workflow structure and individual job characteristics – Task assignment to levels

• Top-Down Approach
• Bottom-Up Approach

– Overall workflow performance estimation

• Based on level characteristics

C D E F A H B G Pietri I., Juve G., Deelman E., Sakellariou R., A Performance Model to Estimate Execution Time

• f Scientific Workflows on the Cloud. In 9th WORKS @ SC14. 2014.

A Performance Model to Estimate Execution Time of Scientific Workflows on the Cloud

7/21

C D E F A H B G

A Performance Model to Estimate Execution Time of Scientific Workflows on the Cloud

• Level-based Estimation Model

– Workflow structure and individual job characteristics – Task assignment to levels

• Top-Down Approach
• Bottom-Up Approach

– Overall workflow performance estimation

• Based on level characteristics

7/21 Pietri I., Juve G., Deelman E., Sakellariou R., A Performance Model to Estimate Execution Time

• f Scientific Workflows on the Cloud. In 9th WORKS @ SC14. 2014.

C D E F A H B G

A Performance Model to Estimate Execution Time of Scientific Workflows on the Cloud

• Level-based Estimation Model

– Workflow structure and individual job characteristics – Task assignment to levels

• Top-Down Approach
• Bottom-Up Approach

– Overall workflow performance estimation

• Based on level characteristics

7/21 Pietri I., Juve G., Deelman E., Sakellariou R., A Performance Model to Estimate Execution Time

• f Scientific Workflows on the Cloud. In 9th WORKS @ SC14. 2014.

Performance Modelling for Montage

8/21

• Estimation of makespan for

– Different number of slots – Different CPU frequencies runtimef=[βt*(f/fmax-1)+1]*runtimefmax

• Use of estimated makespan to predict

– User monetary costs – Energy costs

Using the model in practice

9/21

• The user is interested in executing a

workflow within a deadline

• Given an initial schedule*

– Lower the CPU frequency to execute the tasks and achieve energy savings Utilising the slack time

Improving the Schedule for Energy Savings

* an allocation of the tasks to the resources

10/21

• Slack time:

Maximum delay in task execution without exceeding the deadline

− slackTimet=min(spareTimet->s+ slackTimes), s ϵ succt

* For the exit node:

− slackTimetexit=deadline-finishTimetexit − Where spare time is the maximum delay in task execution without affecting the successors start times spareTimet=startTimes-finishTimet-comCostt->s

[1]

Slack Time

[1] R. Sakellariou and H. Zhao, “A low-cost rescheduling policy for efficient mapping of workflows

• n grid systems,” Scientific Programming, vol. 12, no. 4, pp. 253–262, 2004

11/21

Algorithms with DVFS Techniques

P0 P1 P2 2 4 7 6 3 8 5

Example

1

12/21

Algorithms with DVFS Techniques

P0 P1 P2 4 7 6 8 5

Example

Data transfer 6->8 finishes Deadline Execution of task 8 starts Execution

• f task 8

finishes

2 3 1

12/21

Algorithms with DVFS Techniques

P0 P1 P2 2 4 7 6 8 5

Example

Penalty on execution time Start time

• f task 8

is affected

3 1

12/21

• Frequency scaling may not always be

energy-efficient

– Power savings, but longer execution time: runtimef=[βt*(f/fmax-1)+1]*runtimefmax – Idle power to be considered

Energy Savings

runtimefmax timeidle runtimef

Pfmax Pidle Pf

[1] C.H. Hsu, U. Kremer . The design, implementation, and evaluation of a compiler algorithm for cpu energy reduction. ACM SIGPLAN Notices , 38(5):38–48, 2003. [2] M. Etinski, Corbalan, J. Labarta, and M. Valero, “Understanding the future of energy-performance trade-off via DVFS in HPC environments,” JPDC, vol. 72, no. 4, pp. 579–590, 2012.

13/21

• Workflows may consist of heterogeneous tasks

– Different energy vs frequency behaviour – Reducing the frequency of a task may impact

• verall schedule
• Idea

– Start with an initial schedule e.g. HEFT[1] – Apply frequency scaling iteratively

• Based on the energy gain of each task
• Assessing the impact on overall energy consumption

Idea

[1] H. Topcuoglu, S. Hariri, and M.-Y. Wu, “Performance-effective and low-complexity task scheduling for heterogeneous computing,” IEEE TPDS, vol. 13, no. 3, pp. 260–274, 2002 14/21

ESFS (Energy-aware Stepwise Frequency Scaling)

• Stage 1. Start with an initial schedule at maximum

frequency.

• Stage 2. Using the next available frequency (bound):

− 1. Identify the most profitable tasks in terms of energy gain(beyond a threshold). − 2. Update the plan using the lower frequencies for these tasks. − 3. Assess the impact to overall energy consumption (for the whole workflow). − 4. Continue with the procedure as long as energy savings are achieved (stage 2).

Energy-Aware Workflow Scheduling Using Frequency Scaling

• I. Pietri, R. Sakellariou. “Energy-Aware Workflow Scheduling Using Frequency Scaling”. ICPP

Workshops (PASA), 2014. 15/21

SIPHT with 1000 tasks

[1] I. Pietri, R. Sakellariou. “Energy-Aware Workflow Scheduling Using Frequency Scaling”. ICPP Workshops (PASA), 2014. [2] Q. Huang et al. “Enhanced energy-efficient scheduling for parallel applications in cloud,” in Proceedings of the 12th IEEE/ACM CCGrid. IEEE, 2012, pp. 781–786.

16/21

• We proposed an energy-efficient approach that

selects the CPU frequency for each task.

• In practice, we may need to choose a CPU

frequency per resource (or core). – CPU provisioning charged based on frequency

• Linear relation between frequency and price

e.g. CloudSigma[1] and ElasticHosts[2]

– User: minimize the cost within a deadline

Selecting CPU Frequencies Per Resource

[1] https://www.cloudsigma.com/ [2] http://www.elastichosts.com/ 17/21

CSFS (Cost-based Stepwise Frequency Selection)

• Stage 1. Start with a schedule at maximum frequency.
• Stage 2. Using the next available frequency:

− 1. Select the resources with cost savings for the new frequency. − 2. Update the plan using the chosen frequency for these resources. − 3. Accept the new plan if costs savings and continue with the same procedure (go to 1). − 4. Otherwise terminate.

Cost-efficient Provisioning of Cloud Resources Priced by CPU

• I. Pietri and R. Sakellariou. Cost-Efficient Provisioning of Cloud Resources Priced by

CPU frequency. Best poster award at UCC2014, 2014. 18/21

Montage with 1000 tasks

19/21

Pricing in practice

• How to use pricing to motivate users for

energy-efficient scheduling in practice?

− Trade-off between

• Energy savings for the provider
• Minimization of user cost
• D. Lucanin, I. Pietri, I. Brandic and R. Sakellariou. A Cloud Controller for Performance-Based
• Pricing. In IEEE Cloud 2015, to appear.

20/21

A Cloud Controller for Performance-based Pricing

• D. Lucanin, I. Pietri, I. Brandic and R. Sakellariou. A Cloud Controller for Performance-based
• Pricing. In IEEE Cloud 2015, to appear.
• Motivation

– The impact on application performance from frequency reduction may vary depending on the application characteristics

• Perceived-performance pricing

– CPU provisioning charged based on impact on performance

• CPU frequency used
• VM CPU-boundedness
• Two-stage cloud controller

– Allocation of VMs to PMs – CPU frequency scaling when energy savings exceed the revenue losses

21/21

Thank you for your time!

Cost-Efficient Resource Management for Scientific Workflows on the Cloud