A Hybrid Scheduling Approach for Scalable Heterogeneous...

A Hybrid Scheduling Approach for

Scalable Heterogeneous Hadoop Systems

Authors:

Aysan Rasooli

Douglas G. Down

McMaster University

November 2012

Agenda

McMaster University 2

• Introducing the Hadoop System

• Heterogeneity and Scalability in Hadoop

• Performance Issues of Existing Hadoop Schedulers

• Proposed Hybrid Scheduling System

• Evaluation

• Conclusion

MapReduce

McMaster University 3

Hadoop System

McMaster University 4

Heterogeneity and Scalability in Hadoop

McMaster University 5

• Cluster

• Workload

• User

Hadoop Schedulers

McMaster University 6

• FIFO

• Fair Sharing

• COSHH

Fair Sharing

McMaster University 7

(Zaharia et al., 2010)

• Group jobs into “pools”

• Assign each pool a guaranteed minimum share

• Divide excess capacity evenly between pools

J1

Job Queue

J2

R4

R3

R2

R1

Fair Sharing

McMaster University 8

• Goal: fast response times for small jobs, guaranteed service

levels for long jobs

• Considers Minimum Share satisfaction, Fairness

Drawbacks:

• Does not take into account locality

• Does not take into account heterogeneity

Classes, Suggested Classes for all

Resources

Execution Time of the New Job on

all Resources

Hadoop System

Task Scheduling Process

Queuing Process

Routing Process

Heartbeat New Job

Selected Jobs

Assigned Tasks

COSHH Scheduler

Hadoop System

Task Scheduling Process

Queuing Process

Routing Process

• Estimation of Job execution time across all resources

• Classify the jobs

• Calculate the best set of suggested job classes for each resource

• Select a job for the current free resource using suggested jobs of the Queuing Process

• Consider the fairness and the minimum share satisfaction in the system

COSHH Scheduler

McMaster University 10

• Considering the heterogeneity in the Hadoop system

• Improves Mean Completion Time

• Considers:

• Minimum Share Satisfaction

• Fairness

• Locality

Performance Issues of Existing Schedulers

McMaster University 11

Problem I. Small Jobs Starvation

FIFO :

J1

Job Queue

J2

R4

R3

R2

R1

Performance Issues of Existing Schedulers

McMaster University 12

Problem II. Sticky Slots

Fair Sharing:

R4

R3

R2

R1

Job Fair

Share Running

Tasks

Job 1 2 2

Job 2 2 2

Job Fair

Share Running

Tasks

Job 1 2 1

Job 2 2 2

Performance Issues of Existing Schedulers

McMaster University 13

Problem III. Resource and Job Mismatch

Storage

Unit

slice 1

slice 2

slice 3

Computation

Unit

slot 1

slot 2

slot 3

slot 4

slot 5

R1

Storage

Unit

slice 1

slice 2

slice 3

slice 4

slice 5

slice 6

Computation

Unit

slot 1

slot 2

R2

File 1:

File 2:

J2

Job Queue

J1

Storage

Unit

slice 1

slice 2

slice 3

Computation

Unit

slot 1

slot 2

slot 3

slot 4

slot 5

Storage

Unit

slice 1

slice 2

slice 3

slice 4

slice 5

slice 6

Computation

Unit

slot 1

slot 2

Problem I. Small Jobs Starvation

McMaster University 14

FIFO :

Problem II. Sticky Slots

McMaster University 15

Fair Sharing :

Problem III. Resource and Job Mismatch

McMaster University 16

COSHH:

Performance Issues of Existing Schedulers

McMaster University 17

Experimental Environment

McMaster University 18

Real Hadoop Workloads

McMaster University 19

(Chen et al., 2011)

Scalability Analysis- Results

Job Number Scalability

McMaster University 20

Scalability Analysis- Results

Resource Number Scalability

McMaster University 21

Scalability Analysis- Hybrid Scheduler

McMaster University 22

Scalability Analysis- Hybrid Scheduler

Job Number Scalability

McMaster University 23

Scalability Analysis- Hybrid Scheduler

Resource Number Scalability

McMaster University 24

Conclusion

McMaster University 25

• Performance Issues of Hadoop Schedulers:

• Small Jobs Starvation

• Sticky Slots

• Resource and Jobs Mismatch

• Propose a Hybrid Hadoop Scheduler

Thanks

McMaster University