Healthy nodes are happy nodes - the Fair Tale of Hadoop Scheduling

With Hadoop being a great sell story for parallel data processing, the infrastructure guys are trying to squeeze the maximum buck out of it by pooling multiple consumers. No wonder, there is competition among Hadoop jobs and you guessed it higher incidence of errors. In this post, we analyze an approach for fair scheduling tasks by determining health of a node.

We begin by describing Delay scheduling introduced first in DelayScheduling: A Simple Technique for Achieving Locality and Fairness in Cluster Scheduling (2010) by Matei Zaharia , Khaled Elmeleegy , Dhruba Borthakur , Scott Shenker , Joydeep Sen Sarma , Ion Stoica  . Thereafter we describe the job scheduling based on Node health degree published in Researchon Job Scheduling Algorithm in Hadoop by Yang Xia, Lei Wang, Qiang Zhao, Gongxuan Zhang.

The justice done by fair scheduling is subject to criticism of being too fair or unfair at times. It has been demonstrated to be suspect to head of line scheduling (allocating the top task in sorted queue to free node) and sticky slots (possibility of being allocated to same node again and again).

Further, it suffers from problem of not being able to preserve data locality, the placement of computation near its input data.  “Locality is crucial for performance in large clusters because

network bisection bandwidth becomes a bottleneck”

Enter delay scheduling, “in which a job waits for a limited amount of time for a scheduling opportunity on a node that has data for it. …
When a node requests a task, if the head-of-line job cannot launch a local task, we skip it and

look at subsequent jobs. However, if a job has been skipped long enough, we start allowing it to launch non-local tasks, to avoid starvation”

Snapshot of Hadoop Fair Scheduler algorithm with Delay scheduling:

The above approach sounds pretty good with the data locality taken care. However, Xia et al did not seem quite happy with it. Their contention has been that what if the nodes are not healthy and error prone? What is the use of fairly scheduling a task on a node which is not fit enough to complete the task?

Xia, Wang, Zhao and Zhang have therefore tried to identity variables that define a node and a task, called characteristic variables.

“The characteristic variables of task describe resource usage and demand for MapReduce task…The characteristic variables of node represent the status and quality of computing resources on a node”

Characteristic Variables
Task	Node
average CPU utilization cusage average I/O utilization Iusage average memory usage Musage	average CPU idle rate Cldle average I/O idle rate Ildle average free memory Mldle

The scheduling algorithm is further outlined below:

“1) When an idle task slot can be used, a pool is chosen according to the Fair scheduler;

2) if the node on which the idle task is located is healthy, then go to 5;

3) if the node on which the idle task is located is sub healthy, characteristic variables of the node are separately multiplied by the reciprocal of Health Degree of node , then go to 5;

4) if the node on which the idle task is located is unhealthy, then go to 8,and set timer on node ,its Health Degree is multiplied by 0.9 every x minutes until it reaches the state of sub healthy;

5) Computing Euclidean distances between characteristic variables of job task and characteristic variables of node, sorting them from small to large;

6) Choosing an appropriate job according to Delay Scheduler in the sorted set;

7) Assigning the task of a chosen job to a task slot to run;

8) Finish”

Take note of step 3 where the sub healthy nodes characteristics are ‘loaded’ to arrive at a fairness determined by health. 

We all know schedulers play a major role in architecture of some of the most complex systems as we discussed in one of our earlier post on multimedia search. A lot of text and code has been written on fair and capacity scheduler. This paper definitely describes a unique approach but further tests on high load multi-use clusters can further demonstrate the usage and possible incorporation in future code base.