<quote who="Leo Sutic"> >> Secondly, you assume that all servers are OK with a long start-up. I, > for >> instance, don't tolerate that my servers take 30s to get back up. It's >> wasting my time. > > In the servers that have acceptable startup times, how much of the > startup time is done loading classes?
Did you mean for servers with unacceptable start-up times? Reading a JAR is comparatively fast, linking the classes with all the virtual tables that makes Java slow, that takes considerable time. Not loading a class is always much faster. >> Thirdly, you assume that all the jars only contains classes that are > being >> used. I beg to differ. As a stupid example, rt.jar, is far from loaded >> when I create a small app. > > rt.jar is loaded by the bootstrap classloader and is thus inaccessible > and will never be scanned. Any class loaded by it will have its > getClass().getClassLoader() == null, so it is unreachable. (Think about > it, java.lang.ClassLoader is in rt.jar. Who loads it?) My point was not that you would scan the rt.jar and load all classes, the point was that only a few classes are loaded, and if you trigger loading all the classes, it will take a LONG time... >> With proper SoII, I foresee a large amount of >> implementation classes that are not loaded. > > Point taken. I am striving in my own framework to strong SoII, and expect several to many implementations for a particular specification. > My assumption isn't that a couple of hundred classes is fine. Suppose > Cocoon consists of a hundred bazillion classes, for argument's sake. > Does Cocoon have an acceptable startup time? I assume yes. OK, how much > of that time is spent loading classes? Double that time and add it to > the startup time. Still acceptable? I don't know how many of all the classes that are on Cocoon's classpath that are actually loaded. I believe far from all, but that is speculation. And I fail to see the relevance. I think you are trying to argue, the attribute scan is a fraction of loading the class, and a fraction of start times are spent on loading classes, so therefor its not a problem. That only works if you assume that you don't trigger a lot of unnecessary classloading . > OK, how about this: Instead of scanning *every* class we can get our > hands on, we limit this process to scanning bundles. I.e. if Avalon gets > an equivalent of the .war file, we will only scan such files. This means > that we limit this to the jars that have a potential to contain Avalon > components. > > Suppose we have a directory structure for jars like this: > > lib/ (not scanned) > | > +-- ext (not scanned) > | > +-- endorsed (not scanned) > | > +-- avalon (scanned) Sounds a lot better to me, or until all classes I use are Avalon components ;o) Niclas --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
