From: "andrew cooke" <andrew@...>

Date: Sun, 5 Mar 2006 10:12:20 -0300 (CLST)

Didn't feel like programming last night after work, so I poked around on
my disk (no internet to surf while staying at the flat in La Serena) and
Acar, Blelloch and Harper.

It was pretty interesting.  An adaptive computation is one that changes to
reflect new inputs.  Think of a spreadsheet.  What they did (back in 2002)
was develop a library in ML that allows simple conversion from an normal
fucntional program to one that is adaptive.  In the paper they give an
example, converting a normal quicksort into an adaptive one.  The normal
code sorts a list; they adaptive code gives a list that changes when the
input changes.  Changing a single value in the input to adaptive code only
takes O(log(n)) time - the whole list is not resorted.

Best of all, the conversion is extremely simple - the adaptive code is
almost identical to the original code.

You can imagine how this works in very vague terms - instead of
constructing the code, you contruct a chain of functions with the
structure of the code, and then trigger them as necessary.

This reminded me of some thinking I did some time ago about data
processing in scientific (astronomical) pipelines.  To understand the
relevance you need to have an idea of how data reduction in astronomy is
changing.

The old way of reducing astronomy data was to work on your own computer,
manipulating the data with a fairly sophisticated set of programs.  These
were typically split into tasks and an astronomer would spend a fair
amount of time running different tasks, tweaking parameter values,
re-running, and generally getting a good idea of how the reduction process
affected the results.

The new way of reducing astronomy data is to use a dedicated pipeline at
or near the telescope.  The pipeline is developed and operated by the
observatory; the astronomer receives processed data as a "product".

Now it's not clear to me why an astronomer cannot run a pipeline locally,
at their own department.  A local pipeline combines the advantages of both
approaches above - it gives a default "product", but also allows tweaking
by the user.

OK, how does this fit in with adaptive computing?  Maybe it's already
obvious - adaptive computing would improve the efficiency of a local
pipeline when the astronomer wants to "tweak" parameters.  Instead of
repeating the reduction of all the data, only those parts that depend on
the changed parameters would need to be re-run.

Of course, that wouldn't work with the pipelines we develop at NOAO, since
they are built using a system that doesn't abstract away the distributed
nature of the system - different processes run on different machines, and
communication is "outside" the language used for implementation (the
technology is basically Fortran 66).

But you could imagine doing the same thing with a system that incorporated
distributed processing inside the language - perhaps something built with
Erlang, for example.  Then adaptive computations could exist across a
network of machines.  That would be amazingly cool.  If any forward
thinking observatory wants such a system, email me... :o)

The paper itself is available in PDF form at
http://www.cs.cmu.edu/~umut/papers/popl02.pdf (or Google for the title).

Andrew