#424 Data decomposition in YaspGrid is not optimal
Metadata
| Property | Value |
|---|---|
| Reported by | Markus Blatt (markus@dr-blatt.de) |
| Reported at | Aug 13, 2008 08:56 |
| Type | Bug Report |
| Version | Git (pre2.4) [autotools] |
| Operating System | Unspecified / All |
| Last edited by | Christian Engwer (christi@conan.iwr.uni-heidelberg.de) |
| Last edited at | Oct 25, 2010 21:10 |
| Closed by | Christian Engwer (christi@conan.iwr.uni-heidelberg.de) |
| Closed at | Oct 25, 2010 21:10 |
| Closed in version | Unknown |
| Resolution | Fixed |
| Comment | by r5323 and r5731 |
Description
Yaspgrid uses Torus::optimize_dims() to compute how to distribute the processors onto the grid.
Unfortunately the resulting distribution is far from optimal in some cases.
Calculating the distribution of 20 processors for a grid of dim=3 and size=[125 125 125] leads to the following mapping
Torus<3>: mapping 20 processes onto 20 1 1 torus.
I would rather expect 5 4 1 instead which would leads to less imbalance (Just 3125 instead of 15625 elements difference between some processors)
With increasing diensions the load imbalances become even more annoying.