MPS sudden performance degradation on measure

[merav-aharoni]

Informations

• Qiskit Aer version: 0.14
• Python version:
• Operating system: Linux

What is the current behavior?

Since Qiskit version 0.14, measurement in MPS is very slow.

Steps to reproduce the problem

import qiskit
n = 25 
circ = qiskit.QuantumCircuit(n, n)
circ.measure(range(n), range(n))
sim = qiskit.Aer.get_backend('qasm_simulator')
res = qiskit.execute(circ, sim, shots=100, backend_options={'method': "statevector"}).result()
print("qasm time " + str(res.time_taken))
res = qiskit.execute(circ, sim, shots=100, backend_options={"method": "matrix_product_state", "fusion_enable":False }).result()
print("mps time " + str(res.time_taken)) 

Results:

qasm time 0.3851964473724365
mps time 32.857303857803345

The run-time depends on the number of shots. When I run 1000 shots, time goes up to 300 sec!

What is the expected behavior?

Up to Qiskit 0.13.0, this circuit ran in less than 1 second.

Suggested solutions

Did anything in Aer change that could cause this problem?
Note that I disabled fusion, to make sure that is not causing the problem.

[merav-aharoni]

This appears to be a bug in parallelization. When I remove the #pragma omp parallel for collapse(2) in the function MPS_Tensor::contract_2_dimensions, the performance on this tests goes down from 18 sec to 0.5 sec. @atilag , can you check what changed in the parallelization? adding also @yaelbh to this thread.

[chriseclectic]

@merav-aharoni I would recommend adding OMP conditions like we use in the statevector and density matrix simulators.

The QubitVector class has the following on all its pragmas:

#pragma omp parallel for if (num_qubits_ > omp_threshold_ && omp_threads_ > 1) num_threads(omp_threads_)

The omp_threshold_ parameter can be set using config options, and a default value should be set based on your profiling of when it is beneficial to use multithreading (at the moment use 14 qubits by default for qubitvector/statevector_state).

[chriseclectic]

Closed by #638