-
Notifications
You must be signed in to change notification settings - Fork 5
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Startup performances #53
Comments
The tool stuff is another possible culprit |
So half of that time is spent trying to find the module version. If you run with |
The biggest offenders:
Total of the biggest offenders: 1332ms |
checkModuleVersionsOrShowSuggestions appears to be 200ms when |
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. |
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. |
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. |
…e module file we need and for default module, do not visit modules and stop at the first source file found
…iled output, then in source If found, we do not check for other versions as this is the most likely candidate, and the fastest too
Actually running it is just 10ms, my bad. |
I've made several improvements and now finding a compiled module or a source module is a lot faster. There's still improvement to be done in |
Note that for some reason, parsing a single module file takes 150ms. I guess from initialisation of AntLR? |
Added related ceylon/ceylon-compiler#1631 for compiler speed for trivial modules. |
One thing which would speed it up a lot was if, for each tool, we generated a class (using an annotation processor) which did the parsing for that tool. That way the main tool would only need to instantiate the relevant class and give it the command line arguments, and all the reflection would be avoided. But that's a lot of work. It has been noted that our tooling is a lot like JBoss forge, and I would have suggested that in the long term we look to use that. But then it turns out that their performance isn't great either |
Do you think alternative reflection libs would help? |
No idea. What, specifically, did you have in mind? |
Nothing, just wondering if something like Scannotation would be faster. |
Hah, it looks like the main issue is that we create lots of models instead of just one ;) |
I don't think Scannotation would help because fundamentally something has to instantiate the tool and call a bunch of setters. AFAICS, Scannotation cannot make invoking those setters any faster. Maybe using method handles would help. But avoiding any kind of reflection would be fastest. |
|
Mmm, well that looks stupid. |
So once I removed a few too many instantiations of |
Out of interest what's the cost of |
…ction with NIO2
And bind it only once
So now, a Hello World runs in 800ms. I'm going to stop optimisation for 1.1, and move this issue to 1.2 when we have more time. |
The whole tool setup seems to be around 140ms now. |
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:
compiled with
javac Hello.java
and run withtime java Hello
In average:
ceylon version:
compiled with
ceylon compile hello
and run withtime ceylon run hello
In average:
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)
The text was updated successfully, but these errors were encountered: