Investigate why sometimes fitting results depend on the platform

[Roman Tolchenov]

In some cases results of fitting depend on which platform it is done. The difference is small, within fitting errors. The examples are: Vesuvio fitting scripts, possibly fitting in CalibrateRectangularDetectors (that's why CalibrateRectangularDetector_Test runs only on rhel6).

[Nick Draper]

Moved to R3.5 at the R3.4 code freeze

[Federico M Pouzols]

We have found that this could be an issue for EnginX algorithms, especially for EnginXCalibrate. We still need to check with EnginX instrument scientists what precision they need in their results. At SNS there's some work being done around FindPeaks and Calibrate algorithms at present which may also be affected by this fitting-platform-dependence issue. I might be looking into this soonish, at least to address some basic questions:

• With EnginX we have noticed that OSX persistently produces fitting results that differ of the order of nearly 1% from other platforms. Is OSX wrong, or the others, or all?

• Is the second fitting (linear) done in EnginXCalibrate the one producing or amplifying errors? Hopefully not, as it is a simple linear fit. Clarify how big errors are after the first and second EnginXFitPeaks operations.

So I'm assigning this to me tentatively. Let's see what we find.

[Roman Tolchenov]

It would be nice if you could find more or less simple test cases.

[Federico M Pouzols]

I think that we have some good examples with EnginX. Some of them in our system test, although not necessarily the simplest ones. I'll try to break it down into smaller pieces.

[Stuart Campbell]

This ticket has been transferred to github issue ​10869