Performance regression in shadow variable processing

[triceo]

There needs to be a "happy path" shortcut for when the DSVs cannot form a loop. Currently, the VRP example shows a significant performance regression. The regression should either be completely eliminated, or brought to low single-digit levels.

[simontiffert]

I tried DSV in one customer project with 1.23.0. No loop detection is necessary.

Non-DSV setup: 1x ShadowVariable + 2x PiggybackShadowVariable with 1x VariableListener implementation.
DSV setup: 3x ShadowVariable with 3x @Shadowsources.

DSV shows a speed decrease of factor 8x-10x (test run 10925/sec vs. 1319/sec) which really slows down the solution and I switched back to non-DSV.

If this is a known behavior in the current version this is just for your information. If you need more information let me know.

[triceo]

Hello @simontiffert - thanks for the information!
Yes, it looks very much like the same regression, although much more significant than what we've seen so far. In our benchmarks, it was around 30 %.

[triceo]

@simontiffert The PR we just merged addresses the known regression in use cases where loops are not possible. In fact, our benchmarks show that the new code is even faster than the variable listeners that came before.

If you're still seeing problems in your use case, we want to know.

[simontiffert]

@triceo I just did a test run on this project to verify and the results are impressive. DSV is now as fast Non-DSV on the main branch (did 2 runs on the main branch).

• DSV main branch: 22357/sec + 21108/sec
• Non-DSV main branch: 21231/sec + 21776/sec
• DSV 1.23.0: 3066/sec
• Non-DSV 1.23.0: 22335/sec

[triceo]

Thank you, Simon, much appreciated!