New feature: parallelisation auto-deployer

[TysonRayJones]

contributors: Richard Meister and Tyson Jones
release: v4

QuEST will automatically choose the optimal parallelisation/deployment mode to simulate a newly created Qureg, at runtime. It will consult the Qureg size and type, and the available hardware and library compilations, in an attempt to choose the optimal parallelisation technique.

• user benefits

  • Users can achieve optimal simulation while ignorant of their platform's hardware facilities. For example, multithreading will be automatically disabled if the Qureg is too small to outweigh the OpenMP initialisation overhead. Or, distribution will be avoided if unnecessary and the networking costs would exceed the parallelisation benefit.
  • Users can constrain some settings (force distribution) while having the remaining settings optimised given the constraints. E.g. if the user forces distribution of 18 qubits between 1024 nodes, but sets multithreading to optional, then their effective number of qubits simulated per-node is only 8 (or 256 amplitudes) and multithreading will be automatically disabled.

• feature benefits

  • This can be used/adapated for an auto-tuner, where QuEST performs benchmarking indicative of the deployed hardware and changes its thresholds for automatically using certain backends, etc.

[TysonRayJones]

Added in 2850e02