Startup performance suggestions

borkdude commented 4 years ago

It currenly takes around 120ms on my Macbook Pro 2019 to load-pod bootleg. Which is pretty good, but maybe there are some ways to improve on it.

(ns mybbscript
  (:require [babashka.pods :as pods]))

(println "loading pod")
(time (pods/load-pod "bootleg")) ;;=> 120ms
(require '[pod.retrogradeorbit.bootleg.enlive :as enlive])

(enlive/at "<p>{{here comes foo}}</p>" [:p]
           (enlive/content "foo")) ;;=> 16ms

These are some possible optimizations:

Move describe message to top-level form, making it a compile time constant. The id is not necessary in the reply.
Pre-serialize the describe message to bencode at compile time and simply use print + flush to write it to stdout.

The above two strategies only seemed to help a little bit.

Since the describe message contains lots of client side code, all of this code is evaluated on startup which likely contributes to a performance penalty. The user doesn't use all of these macros all the time. Maybe it's possible to find a way to keep the describe message lean and only send macro and function code on demand.

Maybe we could defer loading of pod namespace code to the require of that namespace. This would work for sci, but not for Clojure. Maybe that's an OK trade-off to make, since Clojure startup is usually around 1.5 or more anyway?

borkdude commented 4 years ago

Some more ideas here: https://github.com/babashka/babashka.pods/issues/14

borkdude commented 4 years ago

Since it's not clear if any of these ideas really significantly speed up things, I'll close this for now, but at least we have it documented somewhere.

retrogradeorbit commented 4 years ago

Thanks for going through all this experimentation

borkdude commented 4 years ago

Yes, updated in babashka/babashka.pods#14. It turns out there are significant gains to be made if we split things up per namespace on demand. It's just that with the example I was initially using, it still used the client side heavy namespaces, that's why I didn't see much difference there, since that's where the most time went. When only using markdown for example, things are much faster when splitting things up per namespace.

borkdude commented 4 years ago

I have implemented a new pod operation to optimize startup time when only some namespaces are used. PR coming.

retrogradeorbit commented 4 years ago

load time on linux slashed from 122 msecs to 18 msecs! Thanks heaps! :partying_face:

retrogradeorbit / bootleg

Startup performance suggestions #65