home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.parallel
- Path: sparky!uunet!destroyer!gatech!hubcap!fpst
- From: Douglas Meyer <dougger@ERC.MsState.Edu>
- Subject: Analysis of Parallel Algorithms/Programs (PVM vs Pablo vs ??)
- Message-ID: <1992Aug26.193154.19970@hubcap.clemson.edu>
- Sender: fpst@hubcap.clemson.edu (Steve Stevenson)
- Organization: Clemson University
- Date: Wed, 26 Aug 1992 19:31:54 GMT
- Approved: parallel@hubcap.clemson.edu
- Lines: 91
-
- [We are having a hard time getting this particular item through
- the network. It has been sent two previous times. If you've
- seen this one before and your last name starts with L-N, please
- do a quick follow-up to let me know. Steve]
-
- I'm currently trying to find simulation/analysis tools which will
- allow parallel algorithm/program developers cull through a number of
- solution options without:
-
- 1. having to write codes for real machines
- 2. having to write complete codes (ie. to shorted the culling
- process, it would be preferable to write skeletal codes
- which describe the general communications and computation
- behaviors)
-
- The focus is intended to be a rough, qualitative analysis of
- "skeletal" programs (programs that do not have the mathematical guts,
- but rather approximate the guts with a library function call which
- simply indicates how many floating point operations would be done in
- critical sections, etc.). These analyses are only required to account
- for very basic, first-order effects, and would probably only permit
- the engineer to eliminate those algorithms that are blatantly bad.
- Metrics of interest would be things like CPU time, "Real" time, and
- average message latency. The focus is on the SPMD paradigm, with
- message passing constrained to a family of "global" or "systematic"
- communications patterns (eg. everybody send to the right, everybody
- broadcast, global sum, etc. such as is found in the Intel iPSC
- library).
-
- As I'm sure you have determined, what we are talking about is modeling
- a parallel machine in such a manner that the algorithm (in the form of
- skeletal code) being analyzed can be "executed" on that model, and
- performance information can then be extracted. I suppose this would
- fall into the broad catagory of synthetic modeling.
-
- So you can see, that these answers are not terribly detailed, and the
- people running this project don't want detailed answers.. just answers
- that provide general guidance.
-
- ------------------ End of description---------------------
-
- I am convinced, through the little bit of reading and research I've
- done in the last week that there are a fair number of tools already
- developed and reasonably mature. PVM and HeNCE clearly are for
- running complete parallel codes; Express is similar and some of its
- analysis abilities are close. As far as I can tell, they are more
- complex (and less speicific) than what is needed. It may be like
- killing an ant with a ICBM. (So motivating the use may be tough.)
-
- Reading a paper on Pablo (The Pablo Performance Analysis Environment -
- University of Illinois/NCSA) pointed out some important facts about
- motivating the user, and providing extensibility and usefulness. I am
- convinced that tools exist which either already do what we need, or
- that our need requires re-thinking. There is a sensibility to the
- Pablo approach which is very appealing. I have a difficult time
- justifying asking algorithm developers to learn to use a tool to write
- very simple parallel algorithms for a "simulation tool" (which would
- take as long to write for a real machine) only to learn qualitatively
- about them (and that only roughly) so that they can eliminate a few
- really awful choices (which would probably be intuitively obvious
- after getting some experience) and then being forced to code all the
- others for the real machine to figure out what's best. It seems
- superfluous right now.
-
- Nonetheless, I'm trying to research the issue, and asking for any help
- you're willing to offer.
-
- - Are you aware of any tools currently available which will
- perform the simple tasks we require?
-
- - What about the Pablo/Picasso stuff - do you think that it has
- merit for us?
-
- Philosophically, can you add you $0.02 to the idea of an tool that
- does this kind of algorithm analysis? Talking with Geoffrey Fox, he
- seemed to indicate that we were looking at a very large investment of
- time and people. (I also am starting to get the impression that a
- project like this has so little "return on investment" compared to
- things like Pablo, HeNCE, Express, etc. that it isn't worth it.)
- Daniel Reed, at NCSA, says that he took the route that we're
- considering for a time, and found it to be fruitless due to growing
- disparity between the source and executable (due to compiler issues
- and languages like Fortran90), loss of utility due to fast turnaround
- of new architectures, etc.
-
- Can you offer us any thoughts?
-
-
- I will compile all the information that I get and post it. It will
- probably be at the end of this week or start of next.
-
-