Preloading and Symfony, a love story

Preloading is a hot topic these days, yet the first versions of PHP 7.4 crashed when it was enabled. Since September with the first experiments (and crash reports), Nicolas is tracking the beast: PHP 7.4.5 finally makes it usable in all contexts and a 75% boost has been measured on a "Hello World" app. Can you expect the same benefits on your apps? Preloading comes with its own challenges. Let's review them and figure how you can get the most out of it for the benefit of your server's efficiency and response time.

First presented at https://online.live.symfony.com/

Transcript

1. & Preloading
   A love story
   

   View Slide

2. Agenda
   •A short history of OPcache
   •Preloading, the theory
   •Preloading, the practice
   •Conclusion
   

   View Slide

3. A short history of OPcache
   

   View Slide

4. Lambda-style before it was hype
   • PHP has a run-and-forget execution model
   • Every request has to reload everything all-the-time
   • Memory leaks and shared state are hard problems
   • This is great for developer efficiency
   • This is great for robustness of apps
   • This makes hosting great
   • This makes scalability easy
   

   View Slide

5. Shared memory (SHM) to the rescue
   • The code is immutable
   • The compilation steps can be cached
   • APC, eAccelerator, Turck MMCache, Zend OpCache FTW
   • But which compilation steps?
   

   View Slide

6. Opcode caches
   • The source code is compiled into a list of instructions,
   the opcodes:
   • ADD, CONCAT, ASSIGN, …, DECLARE_CLASS, DECLARE_FUNCTION,
   DECLARE_INHERITED_CLASS, ADD_INTERFACE, etc.
   • Shared memory holds a hashmap map of (file => opcodes) pairs
   

   View Slide

7. Opcode caches++
   • PHP 5.4: interned strings
   • PHP 5.5: OPcache becomes open-source and builtin
   • PHP 5.6: array declarations stay in shared memory
   • PHP 7.0:
   • all static arrays stay in shared memory
   • post-compilation optimizations
   • PHP 7.1, 7.2, 7.3
   • more post-compilation optimizations
   

   View Slide

8. What remains before 7.4?
   • Check freshness
   • Transfer opcodes into each request’s memory
   • Link parent classes, interfaces and traits
   • Validate signatures, covariance and contravariance
   

   View Slide

9. PHP 7.4
   • Immutable opcodes stay in shared memory
   • And…
   

   View Slide

10. Preloading
    The theory, from https://wiki.php.net/rfc/preload
    

    View Slide

11. Out of scope
    • Bootstraping the kernel or equivalent
    • Loading the services (persistent connections)
    • Computing the response
    

    View Slide

12. Preloading removes
    • Freshness checks
    • Everything related to loading and declaring classes and functions
    • Everything is available out-of-the-box, feels like native
    • (RIP function autoloading, you’ll never have any purpose)
    

    View Slide

13. • BTW, constants are not preloaded
    Constants were not declared
    when preloading was used
    

    View Slide

14. Preloading 101
    • opcache.preload=/some/preloading/script.php in your php.ini
    opcache.preload_user=www-data if running as root (Docker anyone?)
    • Any classes or functions included by this script
    Any classes or functions opcache_compile_file()’ed
    • php-fpm / php -S will load them before serving any requests
    

    View Slide

15. (symbol => opcodes) pairs in SHM
    • Incompatible with several apps on the same FPM server
    • Is really immutable: reboot the server to reload the code
    • opcache_get_status() gives all the info
    • Declarations nested in « if » statements are not preloaded
    • Only fully-resolved classes are preloaded
    (you can ignore the warnings)
    • (oh, it’s not available on Windows)
    

    View Slide

16. Preloading
    The practice
    

    View Slide

17. Which strategy for your preload.php?
    • opcache_compile_file()
    or
    • include() + autoloading?
    

    View Slide

18. opcache_compile_file()?
    

    View Slide

19. https://github.com/symfony/symfony/pull/32032
    

    View Slide

20. https://github.com/composer/composer/issues/7777
    

    View Slide

21. opcache_compile_file()?
    1. Deploy without preloading
    2. Get some HTTP traffic
    3. Dump preload.php from opcache_get_status()
    4. Restart prod with preloading
    5. Automate for every deployment
    • Still loads too many classes (that’s theory)
    • Not practical – requires advanced tooling
    • Not compatible with class_alias()
    

    View Slide

22. Composer to the rescue?
    

    View Slide

23. include() + autoloading?
    • Allows for complex loading strategies
    src/ vendor/
    var/cache/
    

    View Slide

24. The obvious part: preload all services
    List all classes used by services
    (autoloading will cascade to
    parent classes/traits/interfaces)
    

    View Slide

25. Cascade to types used
    on the public API
    (skip protect/private, they lead
    to over-preloading)
    

    View Slide

26. List classes used by
    the implementation
    

    View Slide

27. View Slide

28. List classes used by
    the implementation
    Exclude warmup-time services
    

    View Slide

29. List cache warmer artifacts
    (e.g. compiled Twig templates,
    annotation classes, etc.)
    

    View Slide

30. View Slide

31. include() + autoloading!
    • The configuration drives the useful classes
    • Inline class_exists() declare local sidekicks
    • container.preload/.no_preload for fine tuning
    • Make cache warmers report their artifacts
    

    View Slide

32. • opcache.preload=var/cache/prod/App_KernelProdContainer.preload.php
    • Profit
    • (optionally, fine tune with opcache_get_status() and commit the result)
    • (go patch open-source bundles and libs to maximize their preloading potential)
    include() + autoloading?
    

    View Slide

33. Conclusion
    The practice
    

    View Slide

34. Kudos Dmitry and Nikita
    • include() + autoloading
    • opcache_compile_file()
    • Hello World with Twig and without preloading: 360 req/s
    • with preloading, no cache warmers: 560 req/s (+55%)
    • with preloading, yes cache warmers: 630 req/s (+75%)
    • A typical app spends 15% loading code
    • It’s worth it! (but not worth a complex deployment process)
    PS: there is no memory to share when running on the CLI
    PPS: autoloading + Composer FTW and here to stay
    

    View Slide

35. Thank you!
    See you at Disneyland Paris in December or earlier hopefully!
    

    View Slide