Startup performances

[loicrouchon]

I've noticed that it takes quite some time to startup a ceylon program.

I made the following tests:

Simple hello world program in both java and ceylon

Java version:

public class Hello {
  public static void main(String[] args) {
    System.out.println("hello");
  }
}

compiled with javac Hello.java and run with time java Hello
In average:

real    0m0.060s
user   0m0.050s
sys    0m0.012s

ceylon version:

shared void run(){
  print("hello");
}

compiled with ceylon compile hello and run with time ceylon run hello
In average:

real    0m1.445s
user   0m1.050s
sys    0m0.074s

Ceylon: lastest version from git
Java: openjdk "1.7.0_25"
OS: Ubuntu 13.10 64bits

I don't know if it's those differences are because of JBoss modules initialization, ceylon.language classloading or because of some specific ceylon startup process I'm not aware of.

I'm not really concerned about the fact that it's slower than java, but I would like to know if this can be improved to something more acceptable for command line tools like ceylon.build (below 200ms would be a great start)

[tombentley]

The tool stuff is another possible culprit

[alesj]

Or that model generation?

• https://github.com/ceylon/ceylon-runtime/blob/master/impl/src/main/java/ceylon/modules/jboss/runtime/UtilRegistryTransformer.java#L62

[FroMage]

So half of that time is spent trying to find the module version. If you run with ceylon run hello/1 you will see it takes about 800ms instead of 1500ms.

[FroMage]

The biggest offenders:

• CeylonTool.getTool: 120ms
• checkModuleVersionsOrShowSuggestions: 600ms
• starting JBoss modules: 180ms
• creating JBoss modules class loader: 141ms
• Metamodel init: 91ms
• running the Ceylon code: 200ms

Total of the biggest offenders: 1332ms

[FroMage]

checkModuleVersionsOrShowSuggestions appears to be 200ms when --offline, so the online part is only 400ms

[quintesse]

Yes, the looking up of possible versions comes at a cost, and I agree that it seems pretty high. Most of that probably comes from somewhere deep down in the CMR. For what it does (in most cases just seeing if certain files exists in a certain directories) it seems a bit much.

[quintesse]

Btw, one thing I had been thinking about to prevent the online cost on each invocation is to only query the online repositories if zero or multiple versions were found locally (cached or otherwise). Of course that would make the failure case (no versions exist neither locally nor online) even slower.
(edited for correctness: multiple versions also trigger this situation)

[FroMage]

Yeah, so I'm improving it so that it will use whatever version is already compiled in the current output folder if there's only one, or whatever version is present in source. That's a lot faster.