8u20 with multi-threaded class cache

[Chris Pettitt]

Hannes,

Thanks for the detailed reply - this is very helpful. Your answers clear up my questions for #1 and #2.

For #3, I was hoping we could use some key to retrieve the cached classes that did not involve reloading / re-reading the source scripts on each lookup. jdk.nashorn.internal.runtime.Source is used as the key in jdk.nashorn.internal.objects.Global.findCachedClass. Source tries to avoid the cost of walking the source contents twice, once for hashing (by caching the hash) and once for the equality check by checking the reference. Unfortunately, it *appears* that we're always locked out of this optimization because there is no public way to feed a Source into the NashornScriptEngine.

Thanks,
Chris

________________________________________
From: Hannes Wallnoefer [hannes.wallnoefer at oracle.com]
Sent: Tuesday, March 17, 2015 10:38
To: Chris Pettitt; nashorn-dev at openjdk.java.net; Kunal Cholera
Subject: Re: 8u20 with multi-threaded class cache

Hi Chris, Kunal,

Answers are inlined below.

Am 2015-03-16 um 23:03 schrieb Chris Pettitt:
> Hi folks,
>
> We're looking into the possibility of using the class cache in multi-threaded code, as introduced in [1]. We have a few questions related to this feature:
>
> 1. The article implies that ScriptEngine can be treated as thread-safe - provided we're not using the default context - though the code doesn't state this explicitly. Is this a safe assumption? Are there any other caveats?

This is correct as far as Nashorn is concerned. In the original JDK8
release, using a ScriptEngine with multiple bindings/globals will
compile each script from scratch. We introduced the class caching in
8u20 as our best effort to fit code reuse on the existing ScriptEngine
API. The only caveat I can think of is that code may run slower with
multiple bindings because of callsite polymorphism. Also bear in mind
that this is a relatively new feature, but there seemed to be no problem
using it with your dust scripts (see example below).

> 2. As we need to set the Context for each eval, does this lock us out of using Invocable?

You can actually use this with Invocable.invokeMethod(), passing the
binding as first argument ("thiz" parameter). I've rewritten the
threaded class cache example from the blog post to do this with your
dust benchmark and it seems to work fine:

https://gist.github.com/hns/8f52a620ce36daa3d0ca

I just edited the bench.js file to remove the benchmark loop at the
bottom, otherwise this will run with the script files you sent me. Feel
free to use this as a starting point for your own testing.

It could also work with the other Invocable method taking a "thiz"
parameter, getInterface(Object, Class), but I haven't tested this.

>
> 3. The code for determining if two Sources are the same ultimately falls back to a comparison of the url / name / content of the scripts. Is there a way to eval with a Source to avoid this fallback? It looks like it is not exposed in a public way.

I'm not sure I understand this question. What I use in my example is
jdk.nashorn.api.scripting.URLReader, which is handy to preserve the
source URL in error messages.  I believe Nashorn will still load the
content of the URL to make sure scripts are actually identical.

I hope this helps. Let me know if you have further questions.

Regards,
Hannes

> [1]: https://blogs.oracle.com/nashorn/entry/improving_nashorn_startup_time_using
>
> Thanks,
> Chris