home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!ogicse!uwm.edu!zaphod.mps.ohio-state.edu!darwin.sura.net!gatech!hubcap!fpst
- From: ramani@cse.iitb.ernet.in (K V Ramani)
- Newsgroups: comp.parallel
- Subject: Information wanted on how to estimate communication and
- execution times of a task graph.
- Message-ID: <1992Nov12.131022.2244@hubcap.clemson.edu>
- Date: 11 Nov 92 11:03:11 GMT
- Article-I.D.: hubcap.1992Nov12.131022.2244
- Sender: usenet@cse.iitb.ernet.in (news dummy user)
- Organization: Department of Computer Science,IIT, Bombay
- Lines: 47
- Approved: parallel@hubcap.clemson.edu
- Apparently-To: comp-parallel@uunet.uu.net
-
-
- We are developing a parallelizing compiler which translates
- Fortran-77 programs to a language called Par-Fortran. The target
- language has been designed to run on a network of transputers. The
- front end of the compiler produces a task graph identifying fragments
- of code which can potentially be run in parallel. The nodes of the
- task graph represent statements which should be run sequentially, and
- the edges represent communication between the nodes.
-
- The back end of the compiler has to:
- 1. estimate the execution time of each node and the communication
- along each edge, and use this information to
-
- 2. Partition and schedule the task graph on a given topology of
- transputers, so that the overall execution time is minimized.
-
- Note that, in our context, the program which evaluates the
- execution time has to take to account
-
- a. the time taken to execute the basic instructions of the transputer,
- and,
- b. the code generation strategy of the Par-fortran compiler.
-
- There are essentially two problems. The first concerns
- branches. We are currently assuming that information on the branch
- probabilities are available. We would like to know how to estimate
- branch probabilities. Can these figures be made independent of the
- input?
- The second problem is specific to transputers. The time
- required to execute certain instructions (prod, div etc.) are dependent
- on the values of their operands. These values may not be available at
- compile time. A related problem, which may arise if we use a later
- generation of transputers (T9000-which has multiple functional units
- operating concurrently) is how to estimate execution costs in the
- presence of pipelining and caching.
-
- We would appreciate receiving any relevant information on
- these matters. Any fundamentally different approach towards solving
- these problem would also be useful.
-
- Since access to news is problematic here, we would appreciate if
- responses are sent to us by e-mail.
-
- Thanks in advance.
-
- as@cse.iitb.ernet.in
- ramani@cse.iitb.ernet.in
-