NAME

     Bulk synchronous message passing

DESCRIPTION

     Direct Remote Memory Access ( drma(3)) is a convenient style
     of  programming for BSP computations which can be statically
     analysed in a straightforward way.  It  is  less  convenient
     for  computations  where  the volumes of data being communi-
     cated in supersteps are irregular and  data  dependent,  and
     where the computation to be performed in a superstep depends
     on the quantity and form of data received at  the  start  of
     that  superstep.  A more appropriate style of programming in
     such cases is bulk synchronous message passing (BSMP).


     In BSMP, a non-blocking  send  operation  is  provided  that
     delivers  messages  to  a  system buffer associated with the
     destination process.  The message is guaranteed to be in the
     destination buffer at the beginning of the subsequent super-
     step, and can be accessed by the  destination  process  only
     during  that superstep.  If the message is not accessed dur-
     ing that superstep it is removed from the buffer.  In  keep-
     ing  with  BSP  superstep  semantics, the messages sent to a
     process during a superstep have no implied ordering  at  the
     receiving  end; a destination buffer may therefore be viewed
     as a queue, where the  incoming  messages  are  enqueued  in
     arbitrary  order  and  are  dequeued (accessed) in that same
     order.  Note that although messages are typically identified
     with  tags,  bsplib(3) provides no tag-matching facility for
     the out-of-order access of specific incoming messages.

     In bsplib(3), bulk synchronous message passing is  based  on
     the  idea of two-part messages, a fixed-length part carrying
     tagging information that will help the receiver to interpret
     the  message, and a variable-length part containing the main
     data payload.  We will call the fixed-length portion the tag
     and  the variable-length portion the payload.  The length of
     the tag is required to be fixed during any particular super-
     step,  but  can vary between supersteps.  The buffering mode
     of the BSMP operations is buffered  on source,  buffered  on
     destination .  We note that this buffering classification is
     a semantic description; it does not necessarily describe the
     underlying implementation.

SEE ALSO

     bsp_set_tagsize(3),       bsp_send(3),       bsp_get_tag(3),
     bsp_qsize(3), bsp_move(3), bsp_hpmove(3)

     ``BSPlib: The BSP Programming Library'' Jonathan M. D. Hill,
     Bill  McColl,  Dan  C.  Stefanescu,  Mark W. Goudreau, Kevin
     Lang, Satish B. Rao, , Torsten Suel, Thanasis Tsantilas, and
     Rob  Bisseling.  Parallel  Computing,  to  appear  1998. See
     http://www.bsp-worldwide.org for more details.

BUGS

     Problems  and  bug  reports  should  be  mailed  to  bsplib-
     bugs@comlab.ox.ac.uk

AUTHORS

     The Oxford BSP Toolset implementation of BSPlib was  written
     by Jonathan.Hill@comlab.ox.ac.uk
     http://www.comlab.ox.ac.uk/oucl/people/jonathan.hill.html