Effective warmup strategy for Nashorn - JDK-8u40-b25

Hannes Wallnoefer hannes.wallnoefer at oracle.com
Wed Mar 11 10:06:15 UTC 2015


Hi Kunal,

I'm currently loking at the benchmark scripts you posted.

One approach you should try is to use a single Nashorn ScriptEngine with 
multiple bindings (i.e. JS global objects). This allows Nashorn to share 
compiled classes among multiple threads. It's probably the best 
improvement you can make to startup and warmup time if you're using 
multiple JS environments on multiple threads. On the other hand, it may 
reduce top performance a little bit because of added polymorphism.

I've written about this in my recent Nashorn blog post, it also contains 
demo code for how to set up the script engine for multiple threads:

https://blogs.oracle.com/nashorn/entry/improving_nashorn_startup_time_using

I hope this helps. I'll get back to you when I'm done trying your scripts.

Hannes

Am 2015-03-11 um 02:52 schrieb Kunal Cholera:
> Hello Nahsorn dev team,
>
> Currently our warmup strategy does the following - before our application starts serving request to real users we create 24 Nashorn instances. We are limiting to 24 as we have 24 CPU cores per machine in production. We then load the javascript files into this nashorn instances to setup the context for javascript execution. We then execute each of these javascript 2000 times with mock data to warmup the execution times.
>
> We have observed it takes about 4000 iterations[1] to reach optimal performance. See the graph here that shows 4000 javascript executions on 1 javascript file -
> https://drive.google.com/file/d/0B5PpdCJ7YycqbDJBWGRRVjRsYVU/view?usp=sharing
>
> Three questions
> 1.What is the best way to quickly warm-up to reach optimal performance in fewer iterations ?
>
> 2. Is there a better way to warmup using fixture data as opposed to fake data which we have to maintain ?
>
> 3. Is there a way to clone 1 nashorn instances that is warmed up so that we can avoid doing the same warmup for other nashorn instances ?
>
> [1]
> Steps to reproduce these performance numbers -
> Download this ziphttps://drive.google.com/file/d/0B5PpdCJ7YycqMWlVQ0FsNlozcTA/view?usp=sharing which has the following four files
> 1. bench.js - the benchmarking script
> 2. hp.js - the javascript we want to execute on Nashorn
> 3. hp.json - the data we want to use while executing our script hp.js
> 4. dust-full.js - library needed to execute hp.js script
>
> I am using "1.8.0_40-ea" Java(TM) SE Runtime Environment (build 1.8.0_40-ea-b25 on 8 core redhat linux machine ).
>
> Please advise.
>
> Thanks
> Kunal Cholera



More information about the nashorn-dev mailing list