!!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!!
-------------------------------------------------------------------------------
                                       
                                       
                   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
                                       
      Evaluating Lock Acquirer Prediction and Programming Model in
		   Distributed Shared-Memory Systems

		 	   Ricardo Bianchini 

			University of Rochester


 			   Friday, March 26th
                                       
                                11:00am
                       (Coffee served at 10:45am)

                                       
                         Seminar Room / MCS 135
                                       

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

			       Abstract

  Software-only distributed shared-memory systems (software DSMs) provide
  parallel programmers with the illusion of shared memory on top of
  message-passing hardware at very low cost.  However, many applications
  running on software DSMs suffer high overheads that limit performance.
  In this work we propose and evaluate the Lock Acquirer Prediction (LAP)
  overhead-tolerance technique for software DSMs.  LAP predicts the next
  acquirer of a lock at the time of a lock release operation, based on the
  previous history of lock transfers.  When LAP predicts the next acquirer
  correctly, the shared data at the acquirer can be updated, sometimes
  even before it requests ownership of the lock.  The extent to which
  lock-style primitives are used in parallel applications has a direct
  impact on whether the advantages of LAP can be exploited to the fullest.
  Thus, in order to evaluate LAP we propose a family of Entry
  Consistency-based software DSMs that apply this technique, in which
  successive protocols impose increasingly complex programming models.
  Our experiments with several applications running on an IBM SP2
  multicomputer show that LAP predicts lock acquirers with high accuracy,
  while improving the performance of our software DSMs by as much as 50%.
  Our results also show that the system with the most complex programming
  model outperforms the others by more than 25% on average.  We conclude
  that both LAP and our most complex programming model are features that
  should be provided by future Entry Consistency-based software DSMs.
  
Host: Mark Crovella (crovella@cs.bu.edu)

-------------------------------------------------------------------------------
For colloquium info, including directions, see http://cs-www.bu.edu/colloquium
-------------------------------------------------------------------------------