Proposal: Startup snapshot for optimizing app load time

While this seems interesting, how would it address storing any context information of extension? The code above (in your examples) only stores information about the autoload stack, but anything that would rely on any extension (core or pecl or custom) would have to serialize information that is possibly not even available in userland. Any clue on that? Would it be acceptable to have it just fail/fatal? Marco Pivetta http://twitter.com/Ocramius http://ocramius.github.com/ On 13 April 2016 at 17:55, Lin Yo-An <cornelius.howl@gmail.com> wrote:

On 4/13/16 10:55 AM, Lin Yo-An wrote:

I have no idea how feasible it is for the engine. It may be impossible, or it may be straightforward. However, if it can be done then it would allow for vastly improved architecture for large applications. Many large systems (Drupal being the example I know best) spend a lot of time trying to optimize their bootstrap, loading as few services as possible, adding complexity and indirection to lazy-load services, etc. because that code is executed on every request. That means even a single function call removed along that critical path can add up to a lot of time over the course of a day. If instead we could snapshot a point in time and restart there, we could remove all of that optimization and make the bootstrap process actively load most of the system, then freeze it; then requests come in to a fully booted system and we save a ton of time initializing services on every request. I'd love that. The trick, of course, is that PHP's standard setup model assumes that the environment is created before the code is even loaded. The superglobals, for instance, exist before the first line of userspace code, yet those are the things that will vary per run *after* the snapshot is made. And then there's the question of how to detect and trigger a re-bootstrap, say when code is updated. For that reason, I suspect such a change would require far more than just a few new keywords; there would need to be some sever-level specification of a bootstrap script that is distinct from the user-facing script that the web server executes, and a few thousand edge cases to think about.

Hi, Le 13/04/2016 17:55, Lin Yo-An a écrit :

You are mixing 2 related mechanisms here : code persistence and data persistence. Code persistence (reloading the same code again and again in each request) is achieved through the autoloader. Autoloading, compbined with opcache provides a very fast mechanism to retrieve class definitions. If you measure the time taken by your 'require "vendor/autoload.php"' stance, you should find it to be extremely fast (with opcache on, of course). Data persistence is another question: the need here is to save and restore object instances. Several cache mechanisms exist for this, saving and restoring via serialize/unserialize. The problem with your 'bootstrap' approach (apart from the fact that it is not physically possible in PHP) is that you won't want to snapshot the *whole* environment, because you always want to keep at least some dynamic context. What you actually want to 'snapshot' is a set of well-defined object instances. So, the question, IMO, is to develop a mechanism faster than the current ones to store and retrieve object instances from memory. Regards François

I think this solution is merely a band-aid for a more profound architectural weakness of current PHP setups, where a web server call out to the engine (via embedding or FastCGI) to execute a script, which causes this recurring initialization overhead in the first place. The future is (or should be) servers implemented fully in PHP code (e.g. PHP-PM), just like every other comparable language (Ruby, Python, Java, ...) already does. That also brings many other benefits, such as the ability to handle upgrades to WebSockets in the same codebase, stream request bodies as they're being uploaded, and so forth. And the performance figures that PHP-PM delivers with Symfony validate that approach IMO.

It should be possible to do this for functions and simple classes. Most probably this will require some engine changes, but this is not a big problem for major release. Conditional definition and run-time class inheritance will make significant troubles. Thanks. Dmitry. ________________________________________ From: Lin Yo-An <cornelius.howl@gmail.com> Sent: Wednesday, April 13, 2016 18:55 To: internals@lists.php.net Subject: [PHP-DEV] Proposal: Startup snapshot for optimizing app load time Hi internals, The javascript engine V8 uses a strategy called "startup snapshot" to optimize app load time (see http://v8project.blogspot.tw/2015/09/custom-startup-snapshots.html for more details) The approach used by V8 creates a snapshot from heap, so the V8Context could be reused directly without re-running the bootstraping code. I think running javascript app in the browser is somehow like we run php script for each requests where the app bootstrapping code is required to be executed repeatedly for each request. Currently, our opcache extension can cache the byte codes, but the byte codes still need to be executed every time when bootstrapping an app, and which increases the overhead for each request. For example, a lot of applications use composer to initialize the class loader via the statement below: require "vendor/autoload.php"; The statement above produces thousands of byte code to run, which is to make an app ready. (If you use vld extension to show the opcodes) However, if we can support a simple syntax to describe what is doing "bootstrap", we are able to create a startup snapshot from it. The proposal here want to introduce some new syntax to optimize app load time though the below syntax: bootstrap "vendor/autoload.php"; // caches the context after running this script. Or bootstrap { require "vendor/autoload.php"; // do something else for making app ready }; And of course, we might detect the bootstrap script automatically without creating new syntax and opcodes, but I think it's hard to do because PHP applications can be written by many ways... Questions ======== I don't know if this approach is doable or not. Or maybe we can support this in a separated extension instead of changing the zend core? Do you guys think it's doable? would you like to share your idea for this? Best Regards, Yo-An Lin