| 1 | |
|---|
| 2 | Hej Jakob |
|---|
| 3 | |
|---|
| 4 | (I assume you understand danish) |
|---|
| 5 | |
|---|
| 6 | > I saw your announcement of Pypar on comp.lang.python.announce. In our |
|---|
| 7 | > programs, we are using Konrad Hinsen's Scientific Python module. I |
|---|
| 8 | > have a small technical question. You write that one advantage of your |
|---|
| 9 | > approach is, that one does not need a special version of the Python |
|---|
| 10 | > executable, just the pypar module. In Scientific Python it is |
|---|
| 11 | > necessary to modify Python because MPI requires that MPI_Init is |
|---|
| 12 | > called before the program looks at its command line arguments, and |
|---|
| 13 | > Python needs to look at its command line arguments to figure out which |
|---|
| 14 | > script to run. |
|---|
| 15 | |
|---|
| 16 | Jeg har aldrig rigtigt forstaaet hvorfor det skulle vaere noedvendigt. |
|---|
| 17 | Derfor gik jeg igang med at implementere et interface hvor |
|---|
| 18 | Python oversetteren forbliver uaendret. |
|---|
| 19 | MPI_init har selvfoelgelig ikke adgang til commandline parametre |
|---|
| 20 | - saa det er prisen for min tilgang. |
|---|
| 21 | |
|---|
| 22 | > How do you bypass this problem? I know there are MPI |
|---|
| 23 | > implementations where this is not a problem as MPI only adds extra |
|---|
| 24 | > stuff to the command-line on the non-master nodes, but there are some |
|---|
| 25 | > implementation (MPICH, for example) where the whole command line is |
|---|
| 26 | > replaced with MPI stuff and the command line is only restored when |
|---|
| 27 | > MPI_Init is called. |
|---|
| 28 | |
|---|
| 29 | Det er muligt at pypar vil have problemer med det. |
|---|
| 30 | Jeg har imidlertid testet pypar under LAM/Mpi (linux), |
|---|
| 31 | Native MPI for DEC Alpha og MPICH for Sun enterprise. |
|---|
| 32 | Jeg kunne ikke umiddelbart faa ideen til at virke under MPICH/Linux - maaske |
|---|
| 33 | pga det problem du beskriver. |
|---|
| 34 | |
|---|
| 35 | > I am very interested to hear if you have found an elegant way to solve |
|---|
| 36 | > this, |
|---|
| 37 | |
|---|
| 38 | Jeg har ikke fundet en elegant loesning - maaske snarere en vej udenom: |
|---|
| 39 | Problemet er rent teoretisk for de implementationer jeg bruger |
|---|
| 40 | og kan ignoreres. |
|---|
| 41 | |
|---|
| 42 | |
|---|
| 43 | > as it is slightly annoying to have to have an extra version of |
|---|
| 44 | > the Python executable. |
|---|
| 45 | |
|---|
| 46 | Ja - det var det jeg taenkte. |
|---|
| 47 | |
|---|
| 48 | Hvis du har lyst kan du hjaelpe med at udvikle pypar ideen ? |
|---|
| 49 | |
|---|
| 50 | Mvh |
|---|
| 51 | Ole |
|---|
| 52 | |
|---|
| 53 | ------------------------------------------------------------------- |
|---|
| 54 | Ole Nielsen | Email: Ole.Nielsen@anu.edu.au |
|---|
| 55 | ------------------------------------------------------------------- |
|---|
| 56 | School of Mathematical Sciences| Phone: +61 2 6125 3873 (Direct) |
|---|
| 57 | Australian National University | Fax: +61 2 6125 5549 |
|---|
| 58 | Canberra ACT 0200 | |
|---|
| 59 | Australia | |
|---|
| 60 | ------------------------------------------------------------------- |
|---|
| 61 | |
|---|
