on redundant computation

[Kragen]

On Mon, 14 Dec 1998, Bohn, Christopher A. wrote:
> As a side-note, legend has it that the Manhattan Project had a pool of
> mathematicians who didn't know what they were working on, just that they
> were getting paid to solve equations.  Rather than bogging down the "big
> brains" with working out solutions to equations, the approach was to send
> small sub-subsets of the problem to the mathematician pool (never any
> subsets large enough to indicate what they were for).  Depending on the
> complexity of a particular subset, it would be given to a certain number of
> mathematicians, and if all the solutions agreed, then the solution was
> accepted as a good solution to that subset.  If the solutions disagreed, the
> subset would be further evaluated to figure out why they disagreed.  Kinda
> like what we do with computers today, except that we (often) assume the
> computer's solver is infallible and only have the problem solved once.

This approach is proof not only against accidental errors, but also
against deliberate sabotage -- unless it's coordinated by everyone the
problem is given to.

That means that it might be a reasonable way to manage risks of having
untrusted people compute your answers for you -- the way
distributed.net works, for example.

Another thing -- an Amdahl employee tells me that one of the reasons
mainframes are still popular is that they include massive full-workflow
error-detection capabilities (from the disk through the network into
RAM through the CPU back through RAM back through the network back to
disk), which workstations don't.  Such features could be implemented on
commodity hardware by having the same computations done multiple times
 -- or a "real" computation and a couple of "check" computations on it,
which could result in cheaper hardware and more complex software.

-- 
<kragen@pobox.com>       Kragen Sitaker     <http://www.pobox.com/~kragen/>
Silence may not be golden, but at least it's quiet.  Don't speak unless you
can improve the silence.  I have often regretted my speech, never my silence.
-- Adam Rifkin, <adam@cs.caltech.edu>