thinking about virtual machines

Alaric B. Williams
Mon, 28 Apr 1997 21:16:35 +0000

On 28 Apr 97 at 12:07, wrote:

> For instance what about clusters of MISC computers?
> They'd overpower any stupid CISC/RISC bloatcessor for epsilon of the price.
> but if they have to emulate a CISC/RISC bloatcessor, they will suck.

Hmmm. The problem of such chips is the large memory bandwidth
required to get anything done. I designed a one-instruction chip
once that would be REALLY fast, only modern RAM could never pipe
code into it fast enough (64 bit instructions). Speed right now
depends a lot on how much work is done from each instruction byte.
Viva CISC! Viva CISC!

> Have you really studied the cost of the various compiler passes,
> and wouldn't keeping things high-level (including the generic
> high-level code analyses) keep things much more portable and efficient?

Agreed there; the original point about high level VMs still stands, for
EXACTLY (almost) my reason against MISC :-)

> [Again, using FAT precompiled code for the 2% critical routines].
> What about restrict portability to stubborn von Neuman style machines?
> The von Neuman model is clearly showing its limits performance/cost-wise!

Yeah. Forget MISC, how about FPGAs? They're very fast at repeated operations,
only the have a certain setup time; but the FPGA gatelist can be considered
as a single VERY high level instruction which can be applied to a whole
stream of inputs and yield a whole stream of outputs while still holding
internal state!
> Well, my point is more elaborate:
> 1) I'm not convinced at all that your JVM approach is good
>  for a project like lispOS,
>  but I quite prefer lispOS based on JVM than no lispOS,
>  as I don't think it will lose more than an order of magnitude
>  on speed*size performance.

Agreed. Isn't it nice how we agree so much these days, Fare? Sure
beats the old days :-) How are things these days, anyway? Long time
no decent heated argument!

Alaric B. Williams (

   ---<## OpenDOS FAQ ##>---

Plain HTML:

Fancy HTML: