Re: Simulating an assembler?


John.S.Novak@xxxxxxxxx, III <jsn@xxxxxxxxx> writes:
<SNIP>
> Well, in all seriousness... Yes, I did neglect to mention that
> supercomputers run a hell of a lot faster than laptop computers. Blue
> Gene/L clocks in at 135 Tflops, or 135*10^12 floating point operations
> per second. A laptop would be, what, 5 Gflops? (I'm semi-guessing from
> unreliable sources.) So we're actually talking about a speedup of about
> 25,000. Bear in mind, though, that programming Blue Gene/L is nothing
> like programming your laptop in C++ or Haskell, and the comparison here
> is very crude.

I saw that when I checked the "Top 500" site prior to posting (should
have checked what you actually wrote instead...).

> Second, I'm not sure one would want to directly simulate a device
> running at those speeds for a period of time as long as one second. The
> only reason I can think of to do so would be for reliability
> assessments, but I would think there would be better ways to handle
> those. Generally, you'd want to simulate the device performing each
> motion it is capable of, under a wide selection of typical environments.
> If the simulations are stochastic in nature, than the simulations will
> need to be run often enough to form a meaningful ensemble of results.
> But since they operate in the GHz range, a full second of simulation
> seems rather length.
>
> (Electronics simulators face similar issues, trying to capture the
> essence of necessary initial transient conditions, and then transition
> quickly and smoothely to higher level, less physics-intensive but more
> useful system-level information. In the microwave system simulation
> world, this is a Big Unsolved Problem with no good solutions on the
> horizon. I expect this sort of simulation-fusion issue to be
> nightmarishly complex in comparison to microwave systems. Also, fun.)

Agreed, and this would appear to make the full scale simulations
feasible and useful except for:

> Third, though, I really don't know what the scaling factor in software
> like this would be, if we take number of atoms as a measure of the input
> size. Does the software run-time scale linearly with input size?
> Quadratically? Exponentially? What sorts of approximations does one
> need to make in order to get back to linear from wherever it is?

I imagine this will be critical to the simulation and design of
assemblers. It is a similar situation to the electronic simulators you
mentioned. Hopefully it will be possible to leave out e.g. the
coupling between sufficiently distant parts of the assembler. This
could make the run time linear past a certain size.

Perhaps the detailed simulations of small parts can be used to derive
design rules which, if followed, simplify the simulation.

--

John Devereux

.

Relevant Pages

  • Re: Simulating an assembler?
    ... A laptop would be, what, 5 Gflops? ... If the simulations are stochastic in nature, ... In the microwave system simulation ... Does the software run-time scale linearly with input size? ...
    (sci.nanotech)
  • Re: Covariance matrix
    ... Here is my explanation for your ... from simulations. ... probability distribution, of the form: ... to be some linear but unknown functions of n and s? ...
    (sci.stat.math)
Loading