Re: what if "languages" don't respond to 754R ?

[Nick Maclaren <nmm1@xxxxxxxxxxxxx>]

I have asserted that it is impossible to impose a reproducibility
requirement on parallel codes, and I think that it is fair to justify my
statement.  Here is a sketch of a common class of problem that cannot be
made reproducible without serialising it.  There are many others.

There is a large (indefinite) number of task descriptions in an input
queue (it is irrelevant where that is, on disk or in memory).  Each task
takes an unpredictable amount of time to complete, with no certainty
that its distribution has a mean (i.e. it may have a very long tail,
or it may not).

Each task can be handled independently on a single CPU, and delivers a
floating-point number, normally distributed with mean 0 and variance 1.
The program's specification is to return an estimate of the mean as fast
as possible, subject to it it having a 95% confidence bound of no more
than +-0.001.

One obvious implementation is that threads take a task off the queue
as soon as they are ready, update the sample statistics, and the
program terminates as soon as the limit is reached.  In order for this
to return a reproducible result, the order of completion of each task
must be reproducible.

In order to do that, either the relative time taken to perform two
arbitrary tasks must the same on all systems, or the threads must be
forced to complete in a reproducible order.  The former is ridiculous,
and the latter is incompatible with scalability for at least some
distributions of the time taken for a task.



Regards,
Nick Maclaren,
University of Cambridge Computing Service,
New Museums Site, Pembroke Street, Cambridge CB2 3QH, England.
Email:  nmm1@xxxxxxxxx
Tel.:  +44 1223 334761    Fax:  +44 1223 334679