BlueQuartzSoftware / DREAM3D

Data Analysis program and framework for materials science data analytics, based on the managing framework SIMPL framework.
http://dream3d.bluequartz.net
Other
163 stars 76 forks source link

Some instability of 'Find GBCD' (to be confirmed) #508

Open kglowins opened 9 years ago

kglowins commented 9 years ago

I observed that when I run your 'Find GBCD' filter and plot the distribution, I get slightly different values each time I run the filter (for the same input data). Also the GBCD values in resulting HDF files are not the same. Have you observed similar issues?

sdonegan commented 8 years ago

I haven't noticed this issue; I'll run some tests to see if I can reproduce the problem. Were the values varying a large amount?

kglowins commented 8 years ago

For 3 different runs, I got e.g. 1824.26, 1826.72, and 1821.08, so the differences were not that big.

I don't know whether the problem (if reproducible) occurs in the filter for computing GBCDs or in the one for ploting pole figures.

If this helps, when I was writing my filter, I noticed a similar issue. I had a double value for storing the total area of faces, and it seemed to be different each run (for the same input data). This had an impact on normalization, and therefore, the values of the resulting GBCD. The issue was apparent when parallel processing was turned on, and it was a concurrency problem, because parallel threads were modifying this value.

For GBCDs, the differences were small, like those above. However, for grain boundary plane distributions, where all triangles contribute to the distribution (no requirements for misorientations) and the conflicts are more probable, the differences were up to about 15%.

sdonegan commented 8 years ago

Good to know, thanks for the information. We'll run some more thorough tests on our end and also compare the serial output to the parallel output to see if we can narrow down the issue.