posix-spawn: a faster fork+exec for ruby | Hacker News Reader

posix-spawn: a faster fork+exec for ruby

github.com

56 pointstmm112 comments

This is awesome, Aman's one of my favorite rubyists, I can't say I'm surprised to find out he's behind this.

If anyone's curious how processes/threads are represented internally on modern linux kernels, I wrote an article on that here: http://blog.andrewvc.com/fork2-is-dead-long-live-fork-i-mean...

This is a great article. I came across it while hacking on posix-spawn and it was quite useful.

Wow, check out the Linux vs OSX graphs. Anyone care to enlighten me as to why they are that different? The worst OSX fork + exec call is more than 2x faster than the fastest fork + exec on Linux, and over 10x faster when the memory increases past ~300MB.

Given that OSX's isn't dependent on page table size, I'm guessing it just always does copy-on-write... but that's a pretty un-knowledgeable guess.

The reason fork is so slow on Linux is because the default page size is 4k. This means for a 300MB process, 76800 page table entries have to be copied during the fork.

There are ways to enable HugeTBL on linux to increase the page size, which can improve fork performance as well. See http://stackoverflow.com/questions/2731531/faster-forking-of..., http://linuxgazette.net/155/krishnakumar.html and http://sourceforge.net/projects/libhugetlbfs/

Does anyone know how fork is implemented on BSD/OSX and why it doesn't exhibit the same characteristics?

BSD / OS X has native support for superpages.

From version 2.6.38, Linux will get native huge page / superpage support too. http://lwn.net/Articles/423584/

In FreeBSD, it is extent-based, so a 300MB process might actually only have a couple dozen maps - heap, stack, program binary, and whatever dynamically loaded libraries.

I'd be shocked if it wasn't the same in linux.

I'm curious what the applications are where this is an important optimization. Even for the very large Linux processes, it looks like 1/10sec per fork. On OSX, we're talking milliseconds. What sort of application is forking so often that this matters?

If you use a basic process-per-connection model (select/fork/accept), 10 forks per second is not that much. The "classic Unix" Ruby HTTP servers like Unicorn do just that. (Apache 1.3 and at least one 2.x worker module also work this way, but Apache uses some tricks like preforking and recycling workers to make this fast - I'm not sure the Ruby servers do, too.)

Unicorn pre-forks workers upfront. New processes are not created per request.

posix-spawn is also only useful for fork+exec, where the new process starts executing a new binary. This is not the case in unicorn or resque, where the child processes actually need to be copies of the parent.

We're currently using posix-spawn in albino and grit, which execute external commands like `pygmentize` and `git`

Yes, you're right, of course. Sorry.

On github.com, many types of requests can result in one or more calls to git commands. Every ms we can shave there can have a pretty big impact on response time.

Here's the full announcement: https://gist.github.com/e0d36c881e6efc57514b

Reply on news.ycombinator.com