Polarization Correction Test Too Slow

[Owen Arnold]

Part of the ongoing maintenance task. This one is to speedup PolarizationCorrectionTest, which executes in > 4 seconds. It should run in < 2 seconds.

[Owen Arnold]

refs #10029. Combine operations.

The test code looked lean, I therefore profiled the test suite on RHEL6. The major issue seemed to be the number of binary operations occuring. I have reduced these through some simple rearrangement of the calculations and caching of intermediate values. On my OSX enviroment this is reporting a 50% speed up for the offending test (test_run_PA_unity) which is according to the RHEL6 builds a major bottleneck to the suite.

Changeset: e06a9275f9ff4fa3cd1cc67e06788091c8230bc6

[Owen Arnold]

This test suite is very quick on other platforms, but very slow specifically on RHEL6. At the heart of it, i believe this is because of the same issues proved for #9954 where the cost of thread creation is proving to be very much slower than achieved with the newer Linux kernels. As evidence for this, i have used cachegrind to drill down into the slow binary operations and found that 72% of the CE on RHEL6 is attributed to GOMP_Parallel_Start.

[Owen Arnold]

Tester. The tests should still work. Read the information above (CE per call is 6.5 times larger on RHEL6), and look at the callgrind output to back up what I was saying about cost estimate per call.

I have managed to make the RHEL6 tests mildly quicker.

[Michael Wedel]

Merge remote-tracking branch 'origin/feature/10029_PolCorr_test_speedup'

Full changeset: ​49d891e4db00840920636af87c07394d093f7a90

[Stuart Campbell]

This ticket has been transferred to github issue ​10871