This is a paper motivated by the scheduling problem raised in traditional FIFO strategy in data-intensive cluster computing system.

The proposed methodology is designed to get a good tradeoff point between the conflicts of fairness and data locality, which practically improve response time for small jobs by 5x in a multi-user workload and double throughput in an IO-heavy workload.


Problem with Hadoop

  • Data consolidation provided by a shared cluster is highly beneficial

  • When enough groups began using Hadoop, job response times started to suffer due to Hadoop’s FIFO scheduler

HFS (Hadoop Fair Scheduler)

  • To solve the above raised problem, HFS is designed for two main goals:
    • Fair Sharing
      • Divide resources using max-min fiar shaing to achieve statistical multiplexing
    • Data Locality
      • Place computations near their input data, to maximize system throughput
    • To achieve the first goal, a scheduler must reallocate resources between jobs when the number of jobs changes
    • A key design question is what to do with tasks from running jobs when a new job is submitted, in order to gie resources to the new job

    • At high level, two approaches can be taken
      1. Kill running tasks to make room for the new job
        • Killing reallocates resources instantly, gives control over locality for new job
        • But have drawback for wasting the work of killed tasks
      2. Wait for running tasks to finish
        • Waiting doesn’t waste work from killed tasks
        • But can negatively impact fairness
  • The Principal result in this paper
    • An algorithm based on waiting can achieve both high fairness and high data locality
    • First, in large clusters, tasks finish at such a high rate that resources can be reassigned to new jobs on a timescale much smaller than job durations
    • However, a strict implementation of fair sharing compromises locality
    • Because the job to be scheduled next according to fairness might not have data on the nodes that are currently free
  • To resolve it, the fairness is relaxed slightly through delay scheduling
    • A job waits for a ilmited amount of time for a scheduling opportunity on a node that has data for it
    • A very small amount of waiting is enough to bring locality close to 100%
    • Delay scheduling only asks that we sometimes give resources to jobs out of order to improve data locality


Hadoop Distributed File System

  • Job scheduling at Master
    • Default Scheduler runs jobs in FIFO order, with five priority levels
    • When the scheduler receives a heartbeat indicating that a map or reduce slot is free, it scans through jobs in order of priority and submit time to find one with a task of the required type
  • Locality Optimization for Map operation
    • After selecting a job, the scheduler greedily picks the map task in the job with data closest to the slave

Delay Scheduling

Naive Fair Sharing Algorithm

  • A straight forward strategy is to assign free slots to the job with fewest running tasks
    • As long as slots become free quickly enough, the resulting allocation will satisfy max-min fairness
  • Algorithm 1: Naive Fair Sharing
def scheduler(...):
while a heartbeat is received from node n:
if n has a free slot:
sort jobs in increasing order of number of running tasks
for j in jobs:
if j has unlaunched task t with data on n:
launch t on n
elif j has unlaunched task t:
launch t on n
  • Waiting will not have significant impact on job response time if at least one of the following conditions holds:
    1. Many jobs
    2. Small jobs
    3. Long jobs
  • Data locality problems with Naive Method
    • Head-of-line Scheduling
    • Sticky Slot

Delay scheduling

  • Algorithm 2: Fair sharing with Delay Scheduling
def delay_scheduler(...):
initialize j.skipcnt = 0 for all jobs j
while a heartbeat is received from node n:
if n has a free slot:
sort jobs in increasing order of number of running tasks
for j in jobs:
if j has unlaunched task t with data on n:
launch t on n
j.skipcnt = 0
elif j has unlaunched task t:
if j.skipcnt >= D:
launch t on n
j,skipcnt += 1
  • Note: Once a job has been skipped D tiems, we let it launch arbitrarily many non-local tasks without resetting the skipcount. When if manages to launch a local task again, we set its skipcount back to 0.

Analysis of Delay Scheduling

  • Two interesting observations to the key questions:
    1. How much locality improves depending on D?
      • Non-locality decreases exponentially with D
    2. How long a job waits below its fair share to launch a local task?
      • The amount of waiting required to achieve a given level of locality is a fraction of the average task length and decreases linearly with the number of slots per node L
  • Observation 1:
    • The probability that a job finds at least one local task with skipCount threshold D is:
      • Where $p_j = \frac{\mid P_j \mid}{M}$, $P_j$ is the set of nodes that job j has local data left on it, and $M$ is the total number of nodes in the cluster
  • As we can find that the probability decrease exponentially.
    • For example, when $p_j = 0.1$, and $D=10$, we have $prob = 0.65$ => when $D=40$, we have $prob = 0.99$
  • Observation 2:
    • Once a job j reaches the head of the queue, it will wait at most $\frac{D}{S} \cdot T$ seconds before being allowed to launch non-local tasks.
      • Local tasks run faster than non-local tasks up to 2x

How to set D

  • Suppose we wish to achieve locality greater than $\lambda$ for jobs with N tasks on a cluster with M nodes, L slots per node and replication factor R.

Rack Locality

  • A factor that bandwidth per node within a rack is much higher than bandwidth per node between racks makes it valuable to preserve rack locality
    • This can be accomplished by extending Algorithm 2 to give each job two waiting periods

[1]Delay Scheduling: A Simple Technique for Achieving Locality and Fairness in Cluster Scheduling