This is just a guess, but I think the way Dalvik, the JVM of Android, launches an app is similar to the processing launcher I wrote for Openmoko. It starts a service on dbus, then load and run the processing apps by request. This can make the start up time much shorter because basically it runs in the same jvm instance. Something like this:
fork()
if (child) {
load the app java class
run it
}
/* I'm parent */
wait and check if it's okay
So, during system start up, the first instance of Dalvik is created, and all the common java classes are loaded because you need to start the ‘home’ application. Now whenever you fork a new process, the linux kernel will try to be smart and only copy memory pages if the child process modified something in it. (copy on write)
Anyway, we will see after the release of Android source code.
2008/09/30:
So, it turns out I’m right. According to Anatomy & Physiology of an Android, that first instance is called ‘zygote’. Quite self-explaining.
