[HADOOP-14033] Reduce fair call queue lock contention

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.7.0
• Fix Version/s: 2.9.0, 3.0.0-alpha4, 2.8.2
• Component/s: ipc
• Target Version/s: 2.8.1
• Hadoop Flags: Reviewed
• Assignee: Daryn Sharp
• Reporter: Daryn Sharp
• Created: 27/Jan/17 22:08
• Updated: 27/Sep/17 21:18
• Resolved: 09/Feb/17 22:21

Description

Under heavy load the call queue may run dry yet clients experience high latency.

The fcq requires producers and consumers to sync via a shared lock. Polling consumers hold the lock while scanning all sub-queues. Consumers are serialized despite the sub-queues being thread-safe blocking queues. The effect is to cause other producers/consumers to frequently park.

The lock is unfair, so producers/consumers attempt to barge in on the lock. The outnumbered producers tend to remain blocked for an extended time. As load increases and the queues fill, the barging consumers drain the queues faster than the producers can fill it.

Server metrics provide an illusion of healthy throughput, response time, and call queue length due to starvation on the ingress. Often as the load gets worse, the server looks better.

Issue Links

• breaks: HADOOP-14912 FairCallQueue may defer servicing calls