Investigate spring init startup time

[snicoll]

In 1.4.0.M1, time spring hint takes around 3.2 sec while the same command takes 2.7s on 1.3

[snicoll]

I've done more tests today and I confirm the regression between 1.3.x and master. It's more a 10% increase which is still quite bad.

[joshiste]

The spring-boot-cli-1.4.0.BUILD-SNAPSHOT.jar is around 18M where the spring-boot-cli-1.3.3.RELEASE.jar is around 9M - so I guess the jvm takes simply longer because it has to load more....

Did you thought about disabling the bytecode verification with -Xverify:none and limiting tiered compilation with -XX:+TieredCompilation -XX:TieredStopAtLevel=1 both combined brings me from 0.31s to 0.24s for spring hint in 1.4.0.BUILD-SNAPSHOT

[joshiste]

-- deleted -- was rubbish ... forgot the clean...

[philwebb]

I wonder why the size is so much bigger? That's unexpected.

[joshiste]

yeah I forgot the clean.... so it was nonsense...
size is normal...

[wilkinsona]

I spent a bit of time on this yesterday and didn't get very far. Two things I learned:

• Profiling with YourKit (either sampling or tracing) made 1.4 faster than 1.3 so data was useless
• 1.4 loads very slightly fewer classes than 1.3

[philwebb]

My guess is this is related to #4882

[philwebb]

Prototyping some ideas here: https://github.com/philwebb/spring-boot/tree/gh-5260

1.3.x

java -jar  hint  0.42s user 0.06s system 196% cpu 0.241 total
java -jar  hint  0.40s user 0.05s system 200% cpu 0.226 total
java -jar  hint  0.43s user 0.05s system 207% cpu 0.233 total

Before:

java -jar  hint  0.46s user 0.06s system 203% cpu 0.258 total
java -jar  hint  0.47s user 0.07s system 192% cpu 0.281 total
java -jar  hint  0.47s user 0.06s system 200% cpu 0.265 total

After:

java -jar  hint  0.42s user 0.06s system 186% cpu 0.254 total
java -jar  hint  0.42s user 0.05s system 187% cpu 0.252 total
java -jar  hint  0.41s user 0.06s system 184% cpu 0.253 total

[philwebb]

I think performance is roughly back. I've added #5668 with another idea that might improve things even more.

[snicoll]

I've tested this and we're back to the performance of 1.3.x - Thanks!