[HADOOP-10694] Remove synchronized input streams from Writable deserialization

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.9.0, 3.0.0-alpha1
• Component/s: io
• Labels: BB2015-05-TBR
• Hadoop Flags: Reviewed
• Release Note: Remove invisible synchronization primitives from DataInputBuffer
• Assignee: Gopal Vijayaraghavan
• Reporter: Gopal Vijayaraghavan
• Created: 13/Jun/14 19:14
• Updated: 12/May/16 18:25
• Resolved: 11/May/16 00:29

Description

Writable deserialization is slowing down due to a synchronized block within DataInputBuffer$Buffer.

ByteArrayInputStream::read() is synchronized and this shows up as a slow uncontested lock.

Hive ships with its own faster thread-unsafe version with hive.common.io.NonSyncByteArrayInputStream.

The DataInputBuffer and Writable deserialization should not require a lock per readInt()/read().

Issue Links

• relates to: TEZ-2605 [Umbrella] CPU optimizations for hotspots