[KAFKA-13629] Client producer use faster ByteUtils sizeOfXxx algorithm

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 3.2.0
• Component/s: clients, producer
• Assignee: Unassigned
• Reporter: Jason Koch
• Reviewer: Ismael Juma
• Created: 30/Jan/22 18:53
• Updated: 08/Feb/22 05:44
• Resolved: 07/Feb/22 03:44

Description

Kafka Java client producer path uses `tryAppend()` requires an estimated size. The sizeOf algorithm currently uses an iterative while loop which can be improved.

On real system this reduces the producer path CPU by 4%. JMH benchmarking shows the algorithm is approx 3x faster (90 op/ms -> 290op/ms).

Specifically, we can use a prepared table with the following lookup which relies on JVM intrinsic for fast performance, and also avoids any branching:.

–

int leadingZeros = Integer.numberOfLeadingZeros(value);
return LEADING_ZEROS_TO_U_VARINT_SIZE[leadingZeros];

–

Issue Links

• links to: GitHub Pull Request #11721