% ACL presentation for PUMA final review---outline.

Task 5.1.1 Array Communication Library
--------------------------------------

     Puma architecture provides efficient point-to-point comms.
In data-parallel applications many more elaborate comm patterns
recur frequently.  Can be implemented in terms of message-passing,
but better regarded as more global operations: rearrangements
of distributed data (eg, permutation) or simple arithmetic
transformations (eg, prefix) on distributed data.

     ``Array'' in ACL refers to *process* array: library operates within
arbitrary-dimensional arrays of processes.

      Level is somewhere between low-level communication libraries and
libraries providing specialised support for particular applications,
or specific distribution strategies for data arrays.

     -----*-----

Process arrays
--------------
     A process array is a logically multi-dimensional process configuration.
Each direction has a structure called an ``index''
associated with it, eg, $x$, $y$ and $z$.

% figure `array' goes here.

     -----*-----

Interface
---------
     Index structures are passed to communication routines in much the
same way as channels in occam.  For example

     PERM.SHIFT (x, result, data, 3)

shifts the values in `data' 3 sites in the `x' direction, and returns
the shifted values in `result'.

     The index structure also contains the coordinate value for the
local process associated with the index direction:

     my_x = INDEX.coord (x)


     -----*-----

Implementations
---------------
     In the occam/VCR implementation of ACL any two sites connected by
an index have a virtual channel between them.  The data structure
embodying an index contains vectors of these channels.  The channels
are hidden from the user of the communication library.  Thus the same
library interface can be provided in non-channel-oriented communication
environments.

     -----*-----

Selected entry points
---------------------

     PERM.DS            -- arbitrary permutation.
     PERM.SHIFT         -- a particular permutation, see above
     RELATE.DS          -- a generalised permutation

     BROADCAST          -- and variants

     PREFIX.REAL32.SUM  -- and several other prefixes
     GLOB.REAL32.SUM    -- and several other reductions.

     -----*-----

IO
--
     As well as operations for communication *within* a particular
process array, the library has a sophisticated set of routines for
sequential IO.  These allow a process array to talk to an external
process (eg, a master or host process, or other process arrays---perhaps
an array of graphics processors, etc).

     For these purposes a second type of communication structure is
introduced---a logical ``distributed channel''.

     -----*-----

Configuration
-------------
     The communication library was documented and put into a releasable
form earlier this year.  Since then concentrated on dynamic configuration
procedures for process arrays.

     With static process configuration the channel connectivity
of the networks is exposed to the user in the configuration file.
This is a failure of abstraction---the underlying virtual channels
are supposed to be hidden (transparent).

     By configuring ``procedurally''---using the dynamic features
of VCR---have possibility of hiding channels *completely*.  Also,
this run-time configuration is more flexible.

     -----*-----

Dynamic configuration
---------------------

     The implemented ACL primitives for dynamic configuration use the
remote procedure call, dynamic channel creation, and channel-moving
operations provided in VCR.

    One of the most important operations for dynamic configuration is
``channel-copying''.  Given a virtual channel connecting two processes,
the channel-copying operation creates a new, independent, virtual
channel connecting the same processes.

     -----*-----

Recursive spawning of an ``all-to-all'' network
-----------------------------------------------

% figure `life' goes here.