In this post, “virtual machine” refers to process virtual machines, not to
system virtual machines like Qemu or Virtualbox. A process virtual machine is
simply a program which provides a general programming environment — a program
which can be programmed.
Java has an interpreter as well as a virtual machine, and Python has a virtual
machine as well as an interpreter. The reason “virtual machine” is a more
common term in Java and “interpreter” is a more common term in Python has a lot
to do with the major difference between the two languages: static typing
(Java) vs dynamic typing (Python). In this context, “type” refers to
primitive data types — types which suggest the in-memory storage size of
the data. The Java virtual machine has it easy. It requires the programmer to
specify the primitive data type of each variable. This provides sufficient
information for Java bytecode not only to be interpreted and executed by the
Java virtual machine, but even to be compiled into machine instructions.
The Python virtual machine is more complex in the sense that it takes on the
additional task of pausing before the execution of each operation to determine
the primitive data types for each variable or data structure involved in the
operation. Python frees the programmer from thinking in terms of primitive data
types, and allows operations to be expressed at a higher level. The price of
this freedom is performance. “Interpreter” is the preferred term for Python
because it has to pause to inspect data types, and also because the
comparatively concise syntax of dynamically-typed languages is a good fit for
interactive interfaces. There’s no technical barrier to building an interactive
Java interface, but trying to write any statically-typed code interactively
would be tedious, so it just isn’t done that way.
In the Java world, the virtual machine steals the show because it runs programs
written in a language which can actually be compiled into machine instructions,
and the result is speed and resource efficiency. Java bytecode can be executed
by the Java virtual machine with performance approaching that of compiled
programs, relatively speaking. This is due to the presence of primitive data
type information in the bytecode. The Java virtual machine puts Java in a
category of its own:
portable interpreted statically-typed language
The next closest thing is LLVM, but LLVM operates at a different level:
portable interpreted assembly language
The term “bytecode” is used in both Java and Python, but not all bytecode is
created equal. bytecode is just the generic term for intermediate languages
used by compilers/interpreters. Even C compilers like gcc use an intermediate
language (or several) to get the job done. Java bytecode contains
information about primitive data types, whereas Python bytecode does not. In
this respect, the Python (and Bash,Perl,Ruby, etc.) virtual machine truly is
fundamentally slower than the Java virtual machine, or rather, it simply has
more work to do. It is useful to consider what information is contained in
different bytecode formats:
- llvm: cpu registers
- Java: primitive data types
- Python: user-defined types
To draw a real-world analogy: LLVM works with atoms, the Java virtual machine
works with molecules, and The Python virtual machine works with materials.
Since everything must eventually decompose into subatomic particles (real
machine operations), the Python virtual machine has the most complex task.
Intepreters/compilers of statically-typed languages just don’t have the same
baggage that interpreters/compilers of dynamically-typed languages have.
Programmers of statically-typed languages have to take up the slack, for which
the payoff is performance. However, just as all nondeterministic functions are
secretly deterministic, so are all dynamically-typed languages secretly
statically-typed. Performance differences between the two language families
should therefore level out around the time Python changes its name to HAL 9000.
The virtual machines of dynamic languages like Python implement some idealized
logical machine, and don’t necessarily correspond very closely to any real
physical hardware. The Java virtual machine, in contrast, is more similar in
functionality to a classical C compiler, except that instead of emitting
machine instructions, it executes built-in routines. In Python, an integer is
a Python object with a bunch of attributes and methods attached to it. In
Java, an int is a designated number of bits, usually 32. It’s not really a
fair comparison. Python integers should really be compared to the Java
Integer class. Java’s “int” primitive data type can’t be compared to anything in
the Python language, because the Python language simply lacks this layer of
primitives, and so does Python bytecode.
Because Java variables are explicitly typed, one can reasonably expect
something like Jython performance to be in the same ballpark as
cPython. On the other hand, a Java virtual machine implemented in Python
is almost guaranteed to be slower than mud. And don’t expect Ruby, Perl, etc.,
to fare any better. They weren’t designed to do that. They were designed for
“scripting”, which is what programming in a dynamic language is called.
Every operation that takes place in a virtual machine eventually has to hit real hardware. Virtual machines contain pre-compiled routines which are general enough to to execute any combination of logical operations. A virtual machine may not be emitting new machine instructions, but it certainly is executing its own routines over and over in arbirtrarily complex sequences. The Java virtual machine, the Python virtual machine, and all the other general-purpose virtual machines out there are equal in the sense that they can be coaxed into performing any logic you can dream up, but they are different in terms of what tasks they take on, and what tasks they leave to the programmer.
Psyco for Python is not a full Python virtual machine, but a just-in-time
compiler that hijacks the regular Python virtual machine at points it thinks it
can compile a few lines of code — mainly loops where it thinks the primitive
type of some variable will remain constant even if the value is changing with
each iteration. In that case, it can forego some of the incessent type
determination of the regular virtual machine. You have to be a little careful,
though, lest you pull the type out from under Psyco’s feet. Pysco, however,
usually knows to just fall back to the regular virtual machine if it isn’t
completely confident the type won’t change.
The moral of the story is that primitive data type information is really
helpful to a compiler/virtual machine.
Finally, to put it all in perspective consider this: a Python program executed
by a Python interpreter/virtual machine implemented in Java running on a Java
interpreter/virtual machine implemented in LLVM running in a qemu virtual
machine running on an iPhone.