What you could do is load classes, knowing the class file size and see
the effect on available RAM after you exercise the methods for several
minutes.

Average is very bad metric here. A lot of instructions will generate 1-3
x86 opcodes, while some of them will be very heavy (invoke*, monitor*,
throw, anewarray, etc). This heaviness of course depends if they are
inlined or passed to some utility function.

I suppose that most open source jvms allow to dump their generated code,
so you will be able to see it yourself. For very easy exercise I would
suggest tya, for more complicated (but more representative for real
world jits) - orp.

In tya, most opcodes generated 2-4 x86 instructions, 2-8 bytes. Heavier
instructions like invoke, could easily go into 40-50 instruction count
range, with probably 100-150 bytes taken.

Maybe you can tell why do you need this metric ? Probably there are
different metrics you can use which will be more reliable ?

Artur