A fair sharing job scheduler

The default job scheduler in Hadoop has a first-in-first-out queue of jobs for each priority level. The scheduler always assigns task slots to the first job in the highest-level priority queue that is in need of tasks. This makes it difficult to share a MapReduce cluster between users because a large job will starve subsequent jobs in its queue, but at the same time, giving lower priorities to large jobs would cause them to be starved by a stream of higher-priority jobs. Today one solution to this problem is to create separate MapReduce clusters for different user groups with Hadoop On-Demand, but this hurts system utilization because a group's cluster may be mostly idle for long periods of time. HADOOP-3445 also addresses this problem by sharing a cluster between different queues, but still provides only FIFO scheduling within a queue.

This JIRA proposes a job scheduler based on fair sharing. Fair sharing splits up compute time proportionally between jobs that have been submitted, emulating an "ideal" scheduler that gives each job 1/Nth of the available capacity. When there is a single job running, that job receives all the capacity. When other jobs are submitted, tasks slots that free up are assigned to the new jobs, so that everyone gets roughly the same amount of compute time. This lets short jobs finish in reasonable amounts of time while not starving long jobs. This is the type of scheduling used or emulated by operating systems - e.g. the Completely Fair Scheduler in Linux. Fair sharing can also work with job priorities - the priorities are used as weights to determine the fraction of total compute time that a job should get.

In addition, the scheduler will support a way to guarantee capacity for particular jobs or user groups. A job can be marked as belonging to a "pool" using a parameter in the jobconf. An "allocations" file on the JobTracker can assign a minimum allocation to each pool, which is a minimum number of map slots and reduce slots that the pool must be guaranteed to get when it contains jobs. The scheduler will ensure that each pool gets at least its minimum allocation when it contains jobs, but it will use fair sharing to assign any excess capacity, as well as the capacity within each pool. This lets an organization divide a cluster between groups similarly to the job queues in HADOOP-3445.

Implementation Status:

I've implemented this scheduler using a version of the pluggable scheduler API in HADOOP-3412 that works with Hadoop 0.17. The scheduler supports fair sharing, pools, priorities for weighing job shares, and a text-based allocation config file that is reloaded at runtime whenever it has changed to make it possible to change allocations without restarting the cluster. I will also create a patch for trunk that works with the latest interface in the patch submitted for HADOOP-3412.

The actual implementation is simple. To implement fair sharing, the scheduler keeps track of a "deficit" for each job - the difference between the amount of compute time it should have gotten on an ideal scheduler, and the amount of compute time it actually got. This is a measure of how "unfair" we've been to the job. Every few hundred milliseconds, the scheduler updates the deficit of each job by looking at how many tasks each job had running during this interval vs. how many it should have had given its weight and the set of jobs that were running in this period. Whenever a task slot becomes available, it is assigned to the job with the highest deficit - unless there were one or more jobs who were not meeting their pool capacity guarantees, in which case we choose among those "needy" jobs based again on their deficit.

Extensions:

Once we keep track of pools, weights and deficits, we can do a lot of interesting things with a fair scheduler. One feature I will probably add is an option to give brand new jobs a priority boost until they have run for, say, 10 minutes, to reduce response times even further for short jobs such as ad-hoc queries, while still being fair to longer-running jobs. It would also be easy to add a "maximum number of tasks" cap for each job as in HADOOP-2573 (although with priorities and pools, this JIRA reduces the need for such a cap - you can put a job in its own pool to give it a minimum share, and set its priority to VERY_LOW so it never takes excess capacity if there are other jobs in the cluster). Finally, I may implement "hierarchical pools" - the ability for a group to create pools within its pool, so that it can guarantee minimum allocations to various types of jobs but ensure that together, its jobs get capacity equal to at least its full pool.