TechPrufer
Get started
Meta
Meta logoMeta · Feature trail

HHVM JIT Compiler Performance Optimization

2 posts · Feb 2010 – May 2025

What this trail covers

Meta's HHVM JIT compiler performance optimization has evolved with the introduction of Jump-Start, phased rollouts to share profile data, and dedicated web tenants for specialized workloads like GenAI. This includes optimizations for GenAI inference traffic, increasing request timeouts, adjusting thread-pool sizing, and leveraging JIT cache and request warm-up profiles. Shadow traffic is also used to maintain coverage in the steady state. These efforts have led to a 30% improvement in GenAI latency.

All Meta trails
  1. Meta’s Full-stack HHVM optimizations for GenAIThis post details optimizations made to Meta's monolithic web tier, specifically HHVM, to better handle GenAI inference traffic. Key contributions include splitting GenAI traffic into a dedicated WWW tenant, increasing request runtime limits from 30 seconds to accommodate longer GenAI requests, adjusting thread-pool sizing for higher concurrency (approx. 1000 threads on GenAI hosts vs. a couple hundred on normal webservers), leveraging HHVM's Jump-Start feature with profiles from the main web tier, implementing request warm-up to pre-cache configuration and service discovery info, and using request shadowing to ensure steady-state JIT cache coverage. These optimizations resulted in a 30% improvement in GenAI latency.
  2. Boosting the performance of virtual machines with Jump-StartIntroduced Jump-Start, a technique that leverages phased rollouts to share JIT profiling data across HHVM instances. This reduces warm-up overhead by allowing the majority of servers to skip profiling compilation and execution, and enables more thorough profiling for improved steady-state optimizations. Evaluated to reduce warm-up overhead by 54.9% and improve steady-state performance by 5.4%.
  3. Redesigning the HHVM JIT compiler for better performanceThis post details the redesign of the HHVM JIT compiler to incorporate profile-guided optimizations (PGO). The key innovation is a multi-gear system that uses an initial profiling gear to collect block-execution frequencies and type information, which then feeds into a second optimizing gear. This allows for the compilation of larger code regions, eliminating redundant type checks and enabling optimizations like hoisting loop-invariant computations. The redesign has improved CPU usage by approximately 15%.
  4. HHVM support for PHP 7This post announces HHVM's support for PHP 7.0.0, including all major language features like anonymous classes and generator delegation. It details the strategy for maintaining backward compatibility with PHP 5 by continuing to support it alongside PHP 7 and providing INI options to enable PHP 7 behavior for incompatible changes. The post encourages users to try nightly builds and report issues before the 3.11.0 stable release.
  5. Under the hood: Box’s HHVM migrationBox details their successful migration to HHVM for their PHP codebase, highlighting significant reductions in latency and increases in infrastructure capacity. The post outlines the challenges faced, including resolving unexpected incompatibilities, circumventing expected ones, revamping PHP deployments for HHVM's JIT compilation and restart cadence, and managing mixed environments. Specific technical details include the differences in multiprocessing models (MPM) between Apache prefork and HHVM, the behavior of `memory_get_usage` with jemalloc, and the implementation of a new deployment strategy involving three HHVM instances for warm-up, rollbacks, and canary releases.
  7. Inside the HHVM lockdownThis post details the results of HHVM's first performance lockdown, focusing on improving maximum requests per second. The team evaluated their methodology and collaborated with the community on realistic benchmarking configurations. Significant gains were observed, particularly in MediaWiki, with a 19% performance improvement. Changes were made to benchmarking tools and framework configurations.
  8. HHVM adoption newsThis post announces the adoption of HHVM by Box, Etsy, and Wikipedia, highlighting their positive experiences. Box reported halved CPU utilization for their PHP application. Etsy achieved greater throughput and improved performance, leading to reduced server costs and power consumption. Wikipedia saw approximately 2x faster results for editors after previews or edits. This post serves as a news update on the external adoption and success of HHVM.
  9. Inside Wikipedia’s transition to HHVMThis post details Wikipedia's transition to HHVM, highlighting a 45% reduction in median page save time and decreased CPU usage. It provides direct comparison data between PHP 5.3 and HHVM during the migration, showcasing the performance improvements for server-side PHP execution, particularly for content editing operations.
  10. Announcing the Hack TranspilerIntroduced the Hack Transpiler (h2tp) to convert Hack code to PHP, enabling backward compatibility for projects. This tool handles features like collections and short lambda expressions by generating equivalent PHP code and providing a 'hacklib' library. The primary use case is as a final build step for releases, allowing projects to benefit from Hack while supporting existing PHP users.
  11. Hack: a new programming language for HHVMIntroduced Hack, a new gradually typed programming language for HHVM that interoperates seamlessly with PHP. Hack reconciles the fast development cycle of PHP with the discipline of static typing, offering features like generics, nullable types, type aliasing, constraints on type parameters, Collections, lambda expressions, and runtime enforcement of return types and parameter types. Developed an instantaneous type checker that runs as a local server, typically completing checks in under 200 milliseconds. Released Hack and associated code migration tools as open source.
  12. HHVM – Implementing MySQLiThis post details the implementation of the MySQLi extension for HHVM, a long-requested feature. The author describes the preparation phase, including rewriting the ini parser and using HNI (HHVM Native Interface) and docskel.php for extension development. The implementation involved passing through calls to older MySQL equivalents and implementing the procedural interface in PHP. Significant effort was put into testing, including updating the Zend test suite to support parallel execution and addressing issues with ZendParamMode. Key challenges overcome included implementing `mysqli_stmt::bind_param()` and `mysqli_stmt::bind_result()` which required support for variable arguments passed by reference in HNI. The post also discusses the current status of MySQLi support, noting that 182 out of 300 Zend tests pass, and highlights missing pieces such as php.ini interaction, differing error messages, and lack of support for the MySQL Native Driver. Future plans include supporting ZendParamMode for PHP functions in HNI and passing more Zend tests, with a release planned for HHVM 2.5.0.
  13. Speeding up PHP-based development with HipHop VMThis post details the progress of HHVM, a JIT compilation approach for PHP. It announces that HHVM now matches HPHPc's performance for Facebook's codebase, with a 13% reduction in CPU time attributed to full JIT support for member instructions. It also introduces "side exits" to handle rare conditions within tracelets and discusses type prediction for method call return types, which yielded a ~2% reduction in CPU time. The post also highlights the unification of production and development environments, the open-sourcing of HHVM, and efforts to improve compatibility with popular PHP applications.
  14. Meet a Facebook Engineer: Keith AdamsThis post introduces the HipHop Virtual Machine (HHVM) as a just-in-time execution environment for PHP, aiming to combine high performance with the interactivity that makes PHP productive. It addresses the 'Sophie's Choice' developers face between a high-performance compiled environment (HipHop compiler) and a high-productivity interpreted one.
  15. The HipHop Virtual MachineThis post introduces the HipHop Virtual Machine (HHVM), a new PHP execution engine based on the HipHop language runtime. HHVM utilizes a tracelet-based JIT compilation approach, where each tracelet consists of type guards, a body, and linkage to subsequent tracelets. This approach simplifies trace cache management and allows for dynamic translation of PHP code into native machine code, aiming to provide performance closer to statically compiled binaries and improve developer productivity through faster edit-reload cycles.
  16. HipHop for PHP: More Optimizations for Efficient ServersThis post details further optimizations to the HipHop compiler and runtime system, building upon the initial switch to HipHop for all production servers. Key improvements include: reducing serialization/unserialization costs in APC, optimizing JSON encoding, minimizing unnecessary reference counting through compiler and runtime optimizations, creating more compact binary code by consolidating and uninlining parts of the generated code, improving memory allocation with jemalloc for better profiling and optimization, and parallelizing compiler phases (parsing, optimization, code generation) to speed up compilation times. These changes resulted in a 70% increase in traffic served on the same hardware.
  17. HipHop for PHP: six months laterThis post details the progress of HipHop for PHP six months after its release. It highlights a 1.8x speed improvement over the initial release, leading to an average 50% CPU usage reduction at Facebook. It also mentions successful integration with Drupal, MediaWiki, phpBB, and WordPress, with WordPress seeing a 2.7x speed increase. Community contributions include external patches and support for FreeBSD and 32-bit environments. The post also notes that improvements to PHP itself have made it about 10% faster than 5.3.
  18. HipHop for PHP: Move FastThis post introduces HipHop for PHP, a source code transformer that converts PHP into highly optimized C++ for compilation. It significantly reduces CPU usage on web servers (average 50% reduction) by performing static analysis, type inference, and code generation to produce semantically equivalent C++ code. It sacrifices rarely used features like eval() for performance gains. The system includes a code transformer, a reimplementation of PHP's runtime system, and rewrites of common PHP extensions. It also introduces HPHPi, an experimental interpreter for development.