Job Scheduling in MapReduce

1
Job Scheduling in MapReduce

• Jaideep Dhok

jaideep_at_research.iiit.ac.in
2
Hadoop Terminology

• JobTracker, TaskTracker
• HDFS
• Task, Job
• Slot
• Queue

3
MapReduce workflow

• Workers (TaskTrackers) ask for tasks,
• Master (JobTracker) chooses tasks from jobs in a
  queue and assigns them to workers

4
Issues in MapReduce scheduling

• Maps before Reduce
• Data Locality
• Intermediate Data

5
Issues in MapReduce scheduling

• Speculative Execution
• Shared Environments
• Example Facebook, 3200 jobs/day, average job has
  50 maps.
• Preemption
• Job/Task Recovery

6
Data Locality

• Try to allocate data local tasks, if available
• If not, assign a task anyway
• Problem?
• Yes, for small jobs.

7
Shared Environments

• Multiple users share a single cluster
• FIFO clearly not suitable
• Isolation is important
• Multiplexing improves utilization

8
Expectations from a scheduling algorithm

• Fairness
• Avoid starvation
• Maximize throughput
• Minimize response time
• Optimal use of resources

9
Native Scheduler

• FIFO
• Limited support for job priorities
• Memory aware limit physical and virtual memory
  allowed for a task
• Start reduce after all the maps have finished?

10
FAIR Scheduler

• Pick a job that we have been most unfair to
• Similar examples
• Linux CFS process scheduler
• Linux CFQ disk scheduler

11
FAIR Scheduler - Goals

• Improve response time of small jobs
• Guarantee resources for large jobs
• Multiplex job execution

12
FAIR scheduler - Pools

• Not exactly WFQ (Weighted Fair Queuing)
• Create pools of jobs
• Each pool guaranteed a minimum share of slots
• Divide slots equally between jobs in a pool
• Default, one pool per user

13
FAIR Scheduler

• Locality wait
• Weighted share allows job weights and pool
  weights
• Preemption
• FIFO pools
• Limitations?

14
Capacity Scheduler

• Multiple Queues
• Queues are guaranteed a fraction of the capacity
  of the cluster
• Useful in very large clusters.
• Memory management

15
Capacity Scheduler

• Select a queue if
• It needs to reclaim capacity
• Whose ratio of running tasks to guaranteed
  tasks is lowest
• Within a queue, use FIFO
• Reclaim capacity by killing tasks
• User quotas

16
Dynamic Priority Scheduler

• In FAIR, and Capacity schedulers users have to
  negotiate their share with the administrator
• Not feasible when you have lots of users
• Priority depends on lots of factors
• Ex. Conference deadlines

17
Dynamic Priority Scheduler

• Users given a fixed budget
• Users define their spending rate
• Every allocation deducts money from the users
  account

18
Dynamic Priority Scheduler

• Supports Preemption
• Each user has a different queue
• Amount charged not always equal to spending rate
• Effective rate
• Excess allocations
• Below guaranteed allocation

19
Dynamic Priority Scheduler

• Economic approach to job scheduling
• Current limitations
• Currency is virtual, replaceable
• Users overestimate spending rate
• Abuse

20
Dynamic Priority Scheduler

• Possible extensions
• Demand and Supply driven pricing
• Reservation
• Trading virtual currency

21
Open Areas

• Speculative Execution

22
Open Areas

• Context Switching
• Task Migration

23
Open Areas

• Predicting job resource requirements
• Categorizing jobs based on resource requirements

24
Open Areas

• Throttling / Admission Control

25
Open Areas

• Billing and Accounting
• Scalable and accurate resource usage measurement

26
Open Areas

• Federation
• Job migration between Clusters

27
Open Areas

• Enabling the MapReduce Cloud

28
Questions?
29
Thank You