[HDDS-2301] Write path: Reduce read contention in rocksDB

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Done
• Affects Version/s: 0.5.0
• Labels: performance
• Assignee: Supratim Deka
• Reporter: Rajesh Balamohan
• Created: 14/Oct/19 12:48
• Updated: 02/Jun/20 19:14
• Resolved: 02/Jun/20 19:14

Description

Benchmark:
------------
Simple benchmark which creates 100 and 1000s of keys (empty directory) in OM. This is done in a tight loop and multiple threads from client side to add enough load on CPU. Note that intention is to understand the bottlenecks in OM (intentionally avoiding interactions with SCM & DN).

Observation:
-------------
During write path, Ozone checks OMFileRequest.verifyFilesInPath. This internally calls omMetadataManager.getKeyTable().get(dbKeyName) for every write operation. This turns out to be expensive and chokes the write path.

https://github.com/apache/hadoop/blob/trunk/hadoop-ozone/ozone-manager/src/main/java/org/apache/hadoop/ozone/om/request/file/OMDirectoryCreateRequest.java#L155

https://github.com/apache/hadoop/blob/trunk/hadoop-ozone/ozone-manager/src/main/java/org/apache/hadoop/ozone/om/request/file/OMFileRequest.java#L63

In most of the cases, directory creation would be fresh entry. In such cases, it would be good to try with RocksDB::keyMayExist.

 

 

Issue Links

• duplicates: HDDS-2940 mkdir : create key table entries for intermediate directories in the path
• links to: GitHub Pull Request #107