Improve JarLauncher's performance when running as an exploded jar

[philwebb]

Specifically on https://github.com/spring-projects/sts4/tree/master/headless-services/spring-boot-language-server

[philwebb]

@martinlippert Can you add some instructions on the various ways you start that app and what timings you see?

[martinlippert]

I tried three different things to run the app:

• PropertiesLauncher: java -jar 1.7.0.qualifier-spring-boot-language-server-1.7.0-SNAPSHOT.jar
• JarLauncher: unzip, then java org.springframework.boot.loader.JarLauncher
• plain classpath: unzip, then java -cp "BOOT-INF/classes:BOOT-INF/lib/*" org.springframework.ide.vscode.boot.app.BootLanguagServerBootApp

PropertiesLauncher and JarLauncher is roughly about the same startup time (from the output that appears one the console). The plain class path thing is noticeable faster for me.

[martinlippert]

Our goal is to get the startup time of the overall language server below 1000ms in total, which means that we might have about 500ms available for the pure boot app startup (as measured here), all in total JVM up time. At the moment this is at around ~2300ms for the PropertiesLauncher and the JarLauncher, and around ~1600ms for the plain classpath launch. Looks like this will be an important step towards our goal.

[wilkinsona]

#16248 will, I believe, speed up the second case.

[wilkinsona]

It looks like it may have been wishful thinking that #16248 would help here. Interestingly, I've also observed that java -jar is faster than java org.springframework.boot.loader.JarLauncher. This appears to be the case both with 2.2.0.M2 and with a polished version of the changes proposed in #16248.

[wilkinsona]

I take that back.

Having double-checking my results and built and timed things more scientifically, #16248 does help with the exploded case where the launcher is still used. It brings its start time down so it lies somewhere in between java -cp and java -jar.

	Exploded classpath	Exploded launcher	Jar
2.2 snapshots	0.767	0.974	0.934
gh-16248	0.768	0.851	0.926

The numbers above are the average JVM running time for starting a minimal application, not the language server.

There's still significant room for improvement in the exploded launcher case, but with #16248 at least it's now faster than java -jar which is what I would have expected.

[wilkinsona]

I've tried deferring the creation of the JarFile when the launcher's creating the classpath as all it needs is the URL. Unfortunately, it's made almost no difference – maybe 5ms at best.