Using Amazon APerf to go from 50% below to 36% above performance target

This article discusses using Amazon APerf, a performance analysis tool, to improve the performance of a Java Renaissance benchmark called "finagle-http" on AWS Graviton3 instances.

Specifically, the article covers:

• Overview of APerf and its capabilities for capturing and visualizing system metrics
• Initial performance issue where finagle-http benchmark ran 50% slower on Graviton3 instances compared to a reference system
• Phase 1: Addressing system-level issues identified by APerf such as garbage collection pauses, high CPU interrupts, and kernel scheduling inefficiencies
• Phase 2: Tuning CPU core-level metrics like branch prediction misses, instruction TLB misses, and instructions per cycle (IPC) using JVM options and kernel updates
• By separating the load generator and the benchmark onto different CPU sets, branch prediction was further improved, resulting in a 36% higher performance than the reference system
• Conclusion: APerf enabled identifying and addressing performance bottlenecks, achieving a significant performance improvement