Ampere Performance Toolkit (APT)

Definition:

The Ampere Performance Toolkit (APT) is an open-source performance benchmarking and optimization suite developed by Ampere Computing designed to help developers measure, analyze, and tune application performance specifically on AArch64 (ARM64) architecture — particularly on Ampere’s cloud-native processors like Ampere Altra and Ampere Altra Max.

APT is a one-stop performance lab for ARM64 — giving developers the tools to benchmark workloads, identify bottlenecks, and extract maximum performance from Ampere-based cloud infrastructure.

Background — Why APT Exists Most performance tooling was historically built around x86 (Intel/AMD) processors. As ARM64 servers rapidly entered cloud infrastructure (AWS Graviton, Ampere Altra, etc.), developers needed:

Problem: ❌ x86-tuned benchmarks don’t accurately reflect ARM64 behavior ❌ No unified toolkit for cloud-native ARM64 workloads ❌ Hard to compare performance across different processor architectures ❌ Developers manually assembling dozens of separate tools

APT solves this: ✅ ARM64-native benchmarks and profiling ✅ Unified single toolkit, not a patchwork of tools ✅ Designed for cloud-native workloads (web, AI/ML, databases) ✅ Open-source and freely available on GitHub

Example:

Benchmarking a Web API Server Imagine you’re deploying a high-traffic REST API on an Ampere Altra cloud instance and want to validate performance before going to production.

# Clone the toolkit from GitHub
git clone https://github.com/AmpereComputing/ampere-performance-toolkit
cd ampere-performance-toolkit

# Install dependencies (auto-detects ARM64 environment)
./install.sh

# Verify ARM64 detection
apt-info
# Output:
# Architecture : aarch64 ✅
# CPU Model    : Ampere Altra Q80-30
# CPU Cores    : 80
# Memory       : 256 GB
# OS           : Ubuntu 22.04 LTS

# apt-config.yaml — define what to test

workloads:
  - name: web-server-benchmark
    type: http
    tool: wrk
    target: http://localhost:8080/api/users
    duration: 60s
    threads: 16
    connections: 512

  - name: memory-bandwidth
    type: memory
    tool: STREAM
    iterations: 10

  - name: cpu-compute
    type: cpu
    tool: coremark
    threads: 80    # use all Altra cores

  - name: database-queries
    type: database
    tool: sysbench
    db: postgresql
    tables: 10
    table_size: 1000000
    threads: 32
    duration: 120s

report:
  format: [html, json]
  output: ./results/

apt run --config apt-config.yaml

# Console output:
========================================
  Ampere Performance Toolkit v2.1.0
  Target: Ampere Altra Q80-30 (aarch64)
========================================

[1/4] Running: web-server-benchmark
  → wrk -t16 -c512 -d60s http://localhost:8080/api/users
  → Progress: ████████████████████ 100%
  → Result: 142,847 req/s  ✅

[2/4] Running: memory-bandwidth
  → STREAM benchmark (10 iterations)
  → Progress: ████████████████████ 100%
  → Result: 187.4 GB/s     ✅

[3/4] Running: cpu-compute
  → CoreMark (80 threads)
  → Progress: ████████████████████ 100%
  → Result: 384,200 CoreMark ✅

[4/4] Running: database-queries
  → sysbench PostgreSQL (32 threads, 120s)
  → Progress: ████████████████████ 100%
  → Result: 89,432 TPS      ✅

========================================
  All benchmarks complete ✅
  Report: ./results/apt-report-2024-01-15.html
========================================

APT Performance Report — Ampere Altra Q80-30
════════════════════════════════════════════════

📈 Web Server Throughput
  Result:   142,847 req/s
  Baseline: 120,000 req/s  (previous deployment)
  Delta:    +19.0% improvement ✅

💾 Memory Bandwidth
  Result:   187.4 GB/s
  STREAM Triad score — excellent for 80-core config ✅

⚠️  Bottleneck Detected — CPU Cache Misses
  L3 cache miss rate: 8.2%  (threshold: 5.0%) ⚠️
  Recommendation:
  → Consider NUMA-aware memory allocation
  → Review data structure locality in hot paths
  → Tune thread affinity to physical cores

🗄️  Database Throughput
  Result:   89,432 TPS
  Baseline: 91,000 TPS
  Delta:    -1.7% regression ⚠️
  Recommendation:
  → Profile PostgreSQL connection pool settings
  → Review pg_bouncer configuration for ARM64