Yes I could try to measure J9. I am not sure how class caching works. I tried to use it in my PoC and got confused so I resorted to "classic" openjdk.
-- Michele Sciabarra [email protected] ----- Original message ----- From: Jonathan P Springer <[email protected]> To: [email protected] Subject: Re: Proposal for "fixing" the Java Runtime Date: Mon, 7 Jan 2019 22:29:46 +0000 Could we also take a look at our choice of runtime? We moved to OpenJ9 rather than Hotspot because of the support for class caching, however OpenJ9 is not available (yet) on the ARM64 architecture. Since we’ve got a benchmarking structure in place, can we compare and decide? Sent from my iPhone > On Jan 7, 2019, at 4:35 PM, David P Grove <[email protected]> wrote: > > Michele Sciabarra <[email protected]> wrote on 01/07/2019 01:08:25 PM: >> >> I benchmarked the java runtime, here: >> >> https://sciabarracom.github.io/incubator-openwhisk-runtime-actionloop/ >> >> As you can see the openwhisk/javaaction is pretty slow. Not sure >> why... So I did a proof of concept to see how much I can improve. >> Actually, a lot, as you can see the actionloop/java8, bringing the >> runtime on par with the others using action loop. >> >> The actionloop/java8 is only a PoC, it runs the hello.java but I did >> not run the tests. If the proposal is accepted I will work to make >> sue it passes all the tests of the current java runtime. Then I have >> in mind a lot of extensions, supporting java 11, providing the >> ability to compile sources, download dependencies, compile also >> scala and kotlin... >> >> But let do things in order. It the community ok that I contribute an >> actionloop based java runtime? > > Hi Michele, > > No objections to speeding up Java from me :) > > The gain in /run is impressive. I do think we should try to > understand why there is a slowdown in /init and attempt to mitigate it > (since a cold start will do both /init and /run). > > > --dave
