[HDFS-9901] Move disk IO out of the heartbeat thread

• Type: Improvement
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Component/s: datanode
• Assignee: Hua Liu
• Reporter: Hua Liu
• Created: 04/Mar/16 06:26
• Updated: 16/Dec/16 23:18

Description

During heavy disk IO, we noticed hearbeat thread hangs on checkBlock method, which checks the existence and length of a block before spins off a thread to do the actual transferring. In extreme cases, the heartbeat thread hang more than 10 minutes so the namenode marked the datanode as dead and started replicating its blocks, which caused more disk IO on other nodes and can potentially brought them down.

The patch contains two changes:
1. Makes DF asynchronous when monitoring the disk by creating a thread that checks the disk and updates the disk status periodically. When the heartbeat threads generates storage report, it then reads disk usage information from memory so that the heartbeat thread won't get blocked during heavy diskIO.
2. Makes the checks (which required disk accesses) in transferBlock() in DataNode into a separate thread so the heartbeat does not have to wait for this when heartbeating.

Issue Links

• is duplicated by: HDFS-9910 Datanode heartbeats can get blocked by disk in FsDatasetImpl#checkBlock()
• is related to: HDFS-9882 Add heartbeatsTotal in Datanode metrics