[SPARK-18196] Optimise CompactBuffer implementation

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Won't Fix
• Affects Version/s: 1.6.2, 2.0.1
• Component/s: Spark Core
• Assignee: Unassigned
• Reporter: Adam Roberts
• Created: 01/Nov/16 14:40
• Updated: 11/Dec/16 10:33
• Resolved: 11/Dec/16 10:33

Description

This change slightly increases the class footprint (8 bytes on IBM Java, 12 bytes on OpenJDK and Oracle's) but we've observed a 4% performance on PageRank using HiBench large with this change, so a worthy trade-off IMO

This results in a shorter path length for the JIT as a result of less if else statements

Config used on HiBench

spark.executor.memory 25G
spark.driver.memory 4G
spark.serializer org.apache.spark.serializer.KryoSerializer

Issue Links

• links to: [Github] Pull Request #15713 (a-roberts)