Noiseaware CSPLayout pass

[czachow]

Summary

Refactored the class CSPLayout and created a new class derived of CSPLayout which adds noise awareness.

Closes #5052

Details and comments

• refactored the custom recursive backtracking solver into a separate document
• refactored the method run of CSPLayout into different methods
• created a new method called CSPLayoutNoise
• implemented tests for CSPLayoutNoise

[1ucian0]

It could be nice to have some performance/benchmarking data, if it's not too hard to get.

[czachow]

Thanks for your feedback. I took a look into the other layouters and think it is not easy to compare them. Here you find some benchmarking comparison:

As a benchmark I took the problem of simulating the time evolution of an Ising Chain. The calculation was run on ibmq_bogota using three qubits. Three different layouting methods were used and compared to the simulator. At first the classical CSPLayout (dev), the CSPLayoutNoise (dev_ns) as implemented here and a more advanced CSPLayoutAdvancedNoise (dev_avns) as described in slack were used. The circuit is mapped to different physical qubits (0, 1, 2 in case 1 and 2, 3, 4 in the other). The plot shows the probability of measuring the state 111, starting with an initial state of 000 after a certain time. Additionally, the average of five "independent" runs have been taken. This shows that the noise aware versions perform better and are comparable.

compare_csp_layout.pdf

[1ucian0]

For reference, here is the image in the PDF.

What about runtime? Is it significantly slower than the plain CSP method?

[czachow]

Sorry for the delay.
The difference in runtime strongly depends on the number of solutions S to the layout problem. The additional time complexity can be estimated as the product of the number of gates in the circuit G and the number of solutions found by the crb solver S. I have chosen the same problem as above for an example calculation and averaged over 100 runs of the passmanager. The task is to embed a circuit of three qubits which are only interacting as a chain onto the devices.

This first plot shows the runtime for the ibmq_bogota device for the three different algorithms.

This second plot show the runtime for the ibmq_16_melbourne device for the three different algorithms.

Whereas for the bogota device there is only a small increase, the time increases dramatically for the melbourne device. In my opinion, in a typical setup the number of device and circuit qubits should be approximately equal and thus not that many different possibilities S should arise.

[1ucian0]

After some chat with team and OP, we agree with the following:

• Add a parameter solution_limit as a possible extra limit in the maximum amount of solutions that should be considered. solution_limit=None or solution_limit=0 means "all the solutions". solution_limit=1 has exactly the same effect than current CSPLayout.
• Merge CSPLayoutNoise and CSPLayout.

[1ucian0]

before this PR, we need to merge #6263. Therefore, on hold

[HuangJunye]

Just checking, what's the status of this PR? Why isn't it merged? Still waiting for #6263 ?

[czachow]

In my opinion this merge request can be closed. As a follow-on of this work we (@1ucian0, @georgios-ts and I) investigated the vf2 method for layout allocation. With merge requests #6620, #7213 and #7276 this has now become a part of qiskit.

[HuangJunye]

Luciano told me about the VF2 Layout work. Nice! Thanks for closing the PR as well.