Building a PHP 8.4 Lambda Runtime: Performance Analysis

Why Build Your Own Runtime?

Bref is the most popular way to run PHP on AWS Lambda, and for good reason—it’s well-maintained and gets the job done. However, there are situations where building your own runtime makes sense: you need a specific PHP version before Bref supports it, you want particular extensions, or you simply prefer working with standard PHP patterns without additional abstractions.

This post documents a fully functional PHP 8.4 runtime running on ARM64, complete with performance benchmarks. Consider it a clearly documented alternative to Bref for teams who want complete control over their PHP Lambda environment.

Runtime Architecture

Single Lambda layer (~33MB) containing PHP 8.4 for ARM64, minimal bundled libraries (only libonig), and a bootstrap script implementing the Lambda Runtime API. Supports both PSR-15 HTTP handlers and callable functions.

Performance Benchmarks

Benchmarks across multiple memory configurations on ARM64. Handler processes 10 HAL-formatted entity responses.

Warm Invocation Performance (100 samples each)

Pure CPU performance without I/O:

Key findings:

• 256MB+ configurations all average ~2ms
• 128MB is CPU-bound at 4.2ms (2x slower)
• P99 latency matters: 384MB shows 4ms vs 256MB’s 14ms spikes

Cold Start Performance (30 samples each)

Key findings:

• Init time: 132-147ms (relatively consistent)
• Higher memory = slightly slower init (more resources to allocate)
• Minimal benefit beyond 256MB (only 3% faster at 512MB)

Cost Analysis

At 384MB with AWS Lambda ARM64 pricing ($0.0000133334 per GB-second):

Memory Recommendations

256MB - Best cost/performance ($0.0075/1M, 2.2ms avg) 384MB - Lowest P99 latency (4ms vs 256MB’s 14ms spikes)

Avoid 128MB (2x slower) and >512MB (no performance gain). Function uses ~42MB actual memory.

Comparison with Bref

Comparison to Bref’s published benchmarks:

vs. Swoole/RoadRunner

Swoole and RoadRunner keep PHP processes alive between requests. Lambda’s warm containers achieve similar performance (~2ms/request). The difference: you manage 24/7 servers vs. AWS manages infrastructure with pay-per-request and zero idle cost. Lambda wins for variable traffic; Swoole/RoadRunner for consistent high-volume workloads where you’re already managing servers.

Repository: github.com/marcguyer/lambda-php

This post is part of my serverless PHP series.

This post was written collaboratively with AI assistance, practicing the iterative refinement approach it advocates.

Series: serverless-php

1. Building a PHP 8.4 Lambda Runtime: Performance Analysis

See also

• Building Multi-Architecture AWS Lambda Functions with Bref PHP and PostgreSQL
• Mezzio Example: Testing Doctrine Entities and Repositories
• Mezzio Example: Doctrine Entities and Repositories
• Mezzio Example: Psalm Introduction
• Mezzio Example: Functional and Unit Testing