-------------------------------------------------------------------------------
                                       
                                       
                       B O S T O N   U N I V E R S I T Y
                                       
			  Computer Science Department
				       
			      C O L L O Q U I U M
				       
			   Wednesday April 20, 1994
				       
				    3:00pm
                          (Coffee served at 2:30pm)
                                       
                            Seminar Room / MCS 135
                                       

-------------------------------------------------------------------------------


	    Parallel Programming with Continuation-Passing Threads
				       
				   Yuli Zhou
				      MIT

	Scalable parallel architectures are characterized by physically
	distributed memory, therefore  the issue of achieving good load
	balance  must be addressed under   the constraint of maximizing
	communication  locality. We believe that   the data parallel or
	SPMD model has been very  successful  in solving both  problems
	for a  large   class of  applications   with regular data   and
	communication patterns.  On  the  other hand, for  applications
	with irregular patterns, such as those having recursive control
	and data structures,  the   programmer often implement  ad  hoc
	solutions  using message  libraries,   and that   higher  level
        abstractions are inefficient.
				       
	In this talk  we present  a multi-threaded programming   model,
	which     implements   an     abstraction    of    threads   in
	continuation-passing   style  on top   of  the   basic  message
	transport layer of the machine, aim at providing better support
	for  the  latter class  of  applications.  It includes a thread
	language as  an extension to  C, and a runtime system currently
	implemented on CM-5, enbodying  a work-stealing  scheduler. The
	combination of continuation  passing and work-stealing provides
	an ideal   solution  for   a  class  of  dynamic   applications
	including, for example,  combinatorial search  and optimization
	problems. In the  long term, our  goal is to  evolve the thread
	package  to complement SPMD code by   virtue of its support for
	manipulation  of globally shared  data structures, migration of
	computation   thread and  latency  hiding  with the appropriate
        hardware.

        This is joint work with M. Halbherr of ETH Zurich.

        Host: Prof. Abdelsalam Heddaya

-------------------------------------------------------------------------------
    For more information contact Prof. Azer Bestavros <bestavros@cs.bu.edu>
-------------------------------------------------------------------------------