[YARN-6101] Delay scheduling for node resource balance

• Type: Improvement
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Component/s: fairscheduler
• Assignee: Unassigned
• Reporter: He Tianyi
• Created: 17/Jan/17 09:07
• Updated: 22/Feb/17 09:20

Description

We observed that, in today's cluster, usage of Spark has dramatically increased.
This introduced a new issue that CPU/MEM utilization for single node may become imbalanced due to Spark is generally more memory intensive. For example, after a node with capability (48 cores, 192 GB memory) cannot satisfy a (1 core, 2 GB memory) request if current used resource is (20 cores, 191 GB memory), with plenty of total available resource across the whole cluster.
A thought for avoiding the situation is to introduce some strategy during scheduling.
This JIRA proposes a delay-scheduling-alike approach to achieve better balance between different type of resources on each node.
The basic idea is consider dominant resource for each node, and when a scheduling opportunity on a particular node is offered to a resource request, better make sure the allocation is changing dominant resource of the node, or, in worst case, allocate at once when number of offered scheduling opportunities exceeds a certain number.
With YARN SLS and a simulation file with hybrid workload (MR+Spark), the approach improved cluster resource usage by nearly 5%. And after deployed to production, we observed a 8% improvement.