--- Lun 23/11/09, Riccardo Corradini <[email protected]> ha scritto:
Da: Riccardo Corradini <[email protected]>
Oggetto: R: a few more questions about MPI
A: "Jaroslav Hajek" <[email protected]>
Data: Lunedì 23 novembre 2009, 16:08
Dear Jaroslav,
for question number 3 I just wanted to show how everything could be easy with
MPI_Derived Datatypes (they are not so scary for beginners) because you pass a
lot of pieces of information with just one call.
As a matter of the fact I continue to send vectors, so I handle contiguous
blocks of memory.
At the beginning I had the idea of sending a whole octave_value all at the once
so I thought
MPI_Derived Datatypes has to be the best tool.
Please have also have a look at this in
http://www.osl.iu.edu/research/oompi/
There is a nice example with MPI derived datatypes with the classes.
4) No problems with octave standard of documentation, but I have to learn form
you and others.
Also have a look at this
http://en.wikipedia.org/wiki/Message_Passing_Interface#Derived_Datatypes
They
don't say is better. I just preferred it.
I think more experiments on this should be carried out.
1) No problems renaming the package but I would like to run it on open-mpi
because it seems very spread and I like (it sounds simple).
2) I don't know enough well octave but solution C will be very interesting for
me for learn more about octave. On the other hand I guess you would like to use
solution B.
Respectfully
Riccardo Corradini
--- Lun 23/11/09, Jaroslav Hajek <[email protected]> ha scritto:
Da: Jaroslav Hajek <[email protected]>
Oggetto: a few more questions about MPI
A: "octave-forge list" <[email protected]>
Cc: "Riccardo Corradini" <[email protected]>, "Michael Creel"
<[email protected]>
Data: Lunedì 23
novembre 2009, 15:09
hi guys,
I have a few more comments/questions regarding the MPI package.
1. Will just OpenMPI be targeted or shall we adhere to the MPI standard to
support various implementations?
For instance, the latest probe functions access some non-standard fields of
MPI_State and hence require patching to work under Intel MPI.
1a. The answer to 1 is MPI Standard, what about renaming the package to not
suggest it is OpenMPI specific?
2. My current aim for communicators is just tip of the iceberg; there are a
number of MPI types that may need to be wrapped. In the future.
I see three different approaches here:
A: reinterpret-based
reinterpret the MPI type as some of Octave's types (an integer, or character
array) and vice versa on the function boundary
B: handle-based
store global maps of objects as created by MPI and return integer handles.
C: object-based
derive new Octave objects for the MPI types.
A is completely unsafe, so I think we should avoid that.
B has the advantage that various manipulation of the handles (such as storing
in arrays or structs) would be automatically supported (just like the graphics
handles).
C is the most OO approach, however, there are currently some gotchas in
defining new types from oct files; further, Octave is not really good when a
lot of new types are dynamically registered.
3. this is a question for Riccardo: what are the advantages of using
MPI_Datatype for sending, say, real Matrices, over just providing the "count"
argument to MPI_Send?
4. are you OK with using Octave style coding and documentation?
--
RNDr. Jaroslav Hajek
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz
------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Octave-dev mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/octave-dev
