How Box Slashed Latency by 60% with HHVM: A Deep Dive into PHP Migration

Box’s engineering team recently shared their experience of adopting Facebook’s open‑source HHVM (HipHop Virtual Machine) as the exclusive engine for their PHP services, highlighting significant performance gains and the challenges they faced during migration.

PHP at Box

PHP is the core of Box’s backend stack; every service ultimately interacts with PHP web applications, and the product’s core functionality has been built in PHP since inception.

With over 150 active contributors and more than 750,000 lines of continuously growing PHP code, reducing latency became the top priority. Traditional optimizations and refactoring yielded only marginal improvements.

What is HHVM?

HHVM, originally a PHP‑to‑C++ compiler created by Facebook, has evolved into a JIT (just‑in‑time) compiler that supports the full PHP language, including dynamic features like eval. It now replaces the standard PHP interpreter in many production environments.

Box re‑evaluated HHVM after earlier attempts failed, discovering that HHVM could execute a core endpoint more than four times faster than the default interpreter.

Key Migration Challenges

Achieving feature parity: Porting thousands of PHP extensions and ensuring HHVM matches the default interpreter’s behavior required extensive patches and automated PHPUnit testing.

Design differences: HHVM’s multi‑process model (MPM) uses worker threads instead of Apache prefork processes, affecting logging, file handling, and memory reporting. Adjustments to memory_get_usage and jemalloc settings were necessary to prevent memory leaks.

Deployment changes: HHVM needs a warm‑up period for JIT compilation and periodic restarts for optimal performance, prompting a shift from a single Apache instance to three HHVM instances for seamless roll‑outs and roll‑backs.

Testing and monitoring: A combination of automated and manual tests identified runtime incompatibilities, while continuous monitoring of HHVM error logs and comparison with an internal PHP error database ensured no regressions slipped into production.

Operational Benefits

During the migration, server‑side latency dropped to roughly 2/5 of its original value, and front‑end CPU usage fell by about half, effectively doubling capacity without additional hardware. HHVM also introduced features such as Hack language support, advanced profiling tools, and a vibrant community.

The migration was completed with zero downtime thanks to a host‑by‑host rollout controlled via Puppet flags, allowing rapid roll‑backs within minutes.

In summary, moving Box’s PHP runtime to HHVM was a complex but highly rewarding effort that required careful testing, deployment redesign, and extensive compatibility fixes, ultimately delivering substantial latency reductions, CPU savings, and new capabilities.