Gosling on Computing
CowboyRobot writes "ACM Queue has Eric Allman (creator of Sendmail) interviewing James Gosling (creator of Java) and the conversation covers many aspects of computing today, including the state of security, comparisons of languages and OSs, and the future of virtual machines.
'At the lowest level, you have to know that the boundaries around the piece of software are completely known and contained. So, for example, in Java, you can't go outside the bounds of an array. Ever. Period. Turning off array subscripting is not an option.'"
I lost all my respect to Gosling after a clumsy attempt to add generics to Java.
The right way to add generics to Java is a radical modification of JVM (Java Virtual Machine), but Sun didn't want to it. So they made an attempt to add generics to Java language without touching JVM. The result of this attempt is a complex scheme of name mangling (just like C++), and some unnecessary overhead. And such implementation _still_ requires some JVM changes and is incompatible with old JVMs. So now we have an ugly generics in Java and Java 5.0 (rebranded J2SE 1.5) incompatible with previous versions.
In this article he mentions that the idea for the JVM came from the days when he wrote an emulator and found that his emulator actually performed better than the compiled C code. With the most modern JVM's, tests show that their performance is very close to, and often exceeds that of compiled code.
But the problem is that this is done at the expense of the performance of the overall server. UN*X (and other OS's) go to great lengths to make different programs perform and share resources in the most efficient manner. Scheduling, memory management, IO optimization, all that wonderful stuff, that makes a bear like emacs start almost instantaneously even on an old P90.
As is evident from my sig, I've been spending quite a bit of time in the past year tinkering with virtualization (The Linux Vserver in particular). The amazing thing is that all the optimizations of Linux still apply even if you're running two dozen virtual servers on the same machine (the code for emacs is still shared across processes, even in different virtual servers). Except for, sadly, the JVM. This is why you see many VPS hosting providers forbid running Java and sell a separate "Java server" at a much higher price. And we're considering doing the same. In our experience, a typical VPS customer running Apache, sendmail and a few other things uses 200-400MB of virtual memory, much of which is shared, whereas a customer running Tomcat or, even worse, JBoss use 1-3GB of virtual memory, next to none of which is shared. (Note that virtual memory != physical memory, but that's a separate discussion.
Given that virtualization is becoming more and more popular these days (and even Solaris now has "zones", which are same thing as Linux Vserver or FreeBSD jails), I think the folks at Sun need to think about where the JVM fits into a virtual server.