Performance Regression in multithreading

[ranjanan]

I ran my Logistic Regression benchmark for threads. Not sure if this is related to earlier issues or not, but I thought I'd post an issue anyway.
Usually the output would be (on a version of master sometime in January)

Warm up run : 25 iterations!
Time taken by serial implementation = 1.508228316
Time taken by parallel implementation = 0.429550363
Speedup over 4 threads : 3.5111792374390336

Now, with the current threaded build of master, it's this:

Warm up run : 25 iterations!
Time taken by serial implementation = 1.831504341
Time taken by parallel implementation = 42.974081761
Speedup over 4 threads : 0.04261881268774738

[andreasnoack]

I think this is mainly because of #15276 but it might be useful to keep a separate issue to track the consequences for threading.

[JeffBezanson]

This particular case is fixed on master. If it comes up again, the workaround is to declare types on variables used both inside and outside the @threads loop.

[ViralBShah]

@ranjanan Can you verify?

[ranjanan]

Yes, this seems like it's fixed.

[elanmart]

Hi, running the logreg benchmark from above on 8 threads I'm getting a speedup of only 1.5 on the current master. Could this be a new issue or just a duplicate of #17395?