[HDFS-9761] Rebalancer sleeps too long between iterations

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 2.8.0, 3.0.0-alpha1
• Component/s: balancer & mover
• Target Version/s: 2.8.0
• Hadoop Flags: Reviewed
• Assignee: Mingliang Liu
• Reporter: Adrian Bridgett
• Created: 04/Feb/16 22:34
• Updated: 03/Feb/18 07:13
• Resolved: 05/Feb/16 23:44

Description

Whilst reading the code to try and determine why I'm seeing bad performance I think I've spotted an inadvertent move of a sleep() to inside a for loop.

This means that instead of sleeping once per iteration the balance is sleeping once per namenode per iteration. If I've done my maths correctly the sleep time default is 18seconds so this is a nice little speedup for HA clusters.

Commit d31a41c35927f02f2fb40d19380b5df4bb2b6d57 moved the sleep in org/apache/hadoop/hdfs/server/balancer/Balancer.java:run()

Issue Links

• is broken by: HDFS-8890 Allow admin to specify which blockpools the balancer should run on