| 1 |
edhill |
1.1 |
|
| 2 |
|
|
# $Header: /u/gcmpack/MITgcm/pkg/ex3/README.txt,v 1.1 2005/10/14 05:49:42 edhill Exp $ |
| 3 |
|
|
# $Name: $ |
| 4 |
|
|
|
| 5 |
|
|
======= |
| 6 |
|
|
EX3 |
| 7 |
|
|
======= |
| 8 |
|
|
|
| 9 |
|
|
|
| 10 |
|
|
===== WARNING ===== |
| 11 |
|
|
|
| 12 |
|
|
THIS PACKAGE IS *** UNFINISHED *** AND CONTAINS SHARP, JAGGED BITS |
| 13 |
|
|
THAT ARE LIKELY TO LEAVE NASTY CUTS !!! |
| 14 |
|
|
|
| 15 |
|
|
HANDLE IT AT YOUR OWN RISK ! |
| 16 |
|
|
|
| 17 |
|
|
===== WARNING ===== |
| 18 |
|
|
|
| 19 |
|
|
|
| 20 |
|
|
This is the "ex3" package which stands for "EXchange v3" and it has |
| 21 |
|
|
been designed to do everything that exch2 does with the following |
| 22 |
|
|
advantages: |
| 23 |
|
|
|
| 24 |
|
|
0) EX3 works on more general topologies -- not just cubes! |
| 25 |
|
|
|
| 26 |
|
|
1) A simplified topology "encoding" is used that expresses how the |
| 27 |
|
|
tiles are connected to each other only in terms of cell indicies. |
| 28 |
|
|
It allows for all of the topologies within exch2 plus it has the |
| 29 |
|
|
following advantages: |
| 30 |
|
|
|
| 31 |
|
|
a) Only the cell indicies need to be input by a user -- all the |
| 32 |
|
|
"offsets" needed for exchanging quantities at cell corners, |
| 33 |
|
|
cell faces, etc. are calculated by the exchange routines |
| 34 |
|
|
themselves based upon the simpler cell indicies. |
| 35 |
|
|
|
| 36 |
|
|
b) Any number [n=0,1,2,3,4,...] of tiles may be adjacent at |
| 37 |
|
|
corner points with sufficient information supplied for the |
| 38 |
|
|
vorticity path integrals, |
| 39 |
|
|
|
| 40 |
|
|
c) The array-to-buffer and buffer-to-array copying have been |
| 41 |
|
|
almost completely separated from the logic and data structures |
| 42 |
|
|
which determine where the data sources/sinks are located. |
| 43 |
|
|
This should allow the "encoding" to change without altering |
| 44 |
|
|
the copying routines. |
| 45 |
|
|
|
| 46 |
|
|
d) Tiles with "internal" or "overlapping" connections are now |
| 47 |
|
|
easily represented. |
| 48 |
|
|
|
| 49 |
|
|
2) A more efficient exchange algorithm: |
| 50 |
|
|
|
| 51 |
|
|
a) For vector quantities and scalars defined at vorticity points, |
| 52 |
|
|
a simple "N+1" exchange is first performed to fill the values |
| 53 |
|
|
at the sNx+1 and sNy+1 points. A second pass then fills all |
| 54 |
|
|
the remaining points within the exchange regions. This |
| 55 |
|
|
correctly populates the values around the corners for any |
| 56 |
|
|
number of tiles meeting at a corner point. This is also in |
| 57 |
|
|
contrast to the three-pass algorithm required by exch2. |
| 58 |
|
|
|
| 59 |
|
|
b) Vectors are packed into a single buffer, reducing the number |
| 60 |
|
|
of MPI messages. |
| 61 |
|
|
|
| 62 |
|
|
|
| 63 |
|
|
The main disadvantage of the ex3 package is the removal of support for |
| 64 |
|
|
MDSIO "global" files. With the arrival of non-cube topologies, the |
| 65 |
|
|
cube-oriented MDSIO global file option is not practical since it is |
| 66 |
|
|
(1) not a dense encoding of the data and (2) will require a re-working |
| 67 |
|
|
of the existing MDSIO code. |
| 68 |
|
|
|