home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.parallel
- Path: sparky!uunet!gatech!hubcap!fpst
- From: kaminsky-david@CS.YALE.EDU (David Kaminsky)
- Subject: Re: Linda / The Parform
- Message-ID: <1992Dec21.192950.15404@hubcap.clemson.edu>
- Sender: fpst@hubcap.clemson.edu (Steve Stevenson)
- Organization: Yale University Computer Science Dept., New Haven, CT 06520-2158
- References: <1992Dec17.191745.491@ifi.unizh.ch>
- Date: 21 Dec 1992 14:24:32 -0500
- Approved: parallel@hubcap.clemson.edu
- Lines: 57
-
- > We have compared the open domain Linda implementation POSYBL with a
- > Transputer Multicluster, PVM and our platform The Parform. POSYBL turned
- > out to be clearly worse than the other systems. Thus, we concluded, that
- > the tuple space concept is inefficient.
-
- > After time has passed now, we could make SCA-Linda measurements with an
- > explicit heat equation solver on a 2D grid in our network:
-
- I'd like to thank Volker and Clemens for posting this data. It
- is informative to see a clear comparison of different models of
- parallelism. As we've often argued, Linda is as efficient as message
- passing systems. This data supports our view.
-
- > For further information, we refer the reader to our old report, ftp-able
- > from claude.ifi.unizh.ch. An upgraded version will be published in near
- > future.
-
- I would also like to recommend this paper. It contains some
- interesting work on automatic load balancing.
-
- > This code is message passing code and not Linda flavored code. Runtimes
- > of a Linda flavored code were several times longer than the message
- > passing style code.
-
- I think this statement is misleading. I examined the "Linda
- flavored" code. It approaches a domain decomposition problem using
- dynamic tasking. The worker process repeatedly grabs a column
- (or a block of columns), gets the columns bordering it, updates
- the column, and returns the column to tuple space.
-
- The fact that Linda makes it easy to write this sort of
- program doesn't mean that Linda programmers use this approach blindly
- without regard to the problem. In this case, no decent Linda
- programmer would ever consider doing it this way.
-
- The proper way to write the code is to statically allocate
- blocks of the matrix to the workers and have the workers exchange
- border columns at each iteration. This is the "message passing code"
- to which they refer.
-
- Linda is a superset of message passing and, as this data
- demonstrates, no less efficient than message passing systems on
- message passing problems. It can handle both message passing codes
- (as this one clearly is) and codes not easily handled by message
- passing systems (e.g. those containing shared variables).
- Characterizing a code as "linda flavored" or "message passing"
- shortchanges Linda's flexibility.
-
- For more information about Linda applications, I refer the
- reader to "Experience with Linda", Bjornson, et al. (YALEU/DCS/TR-866)
- and Rob Bjornson's Thesis (1992). Both are available from Chris
- Hatchell (hatchell@cs.yale.edu).
-
-
-
-
-
-