“Automagic” parallel GST using MPI

This tutorial demonstrates how to compute GST estimates in parallel using MPI. This requires the mpi4py python package, which must be able to connect to an underlying MPI library (say, openmpi).

We’ll start by setting the stage.

from pygsti.modelpacks import smq1Q_XYI as mp
from pygsti.protocols import ProtocolData, StandardGST
from pygsti.data import simulate_data

exp_design  = mp.create_gst_experiment_design(max_max_length=32)  # type: ignore
mdl_ideal   = mp.target_model()                                   # type: ignore
mdl_datagen = mdl_ideal.depolarize(op_noise=0.1, spam_noise=0.001)

data  = simulate_data(mdl_datagen, exp_design.all_circuits_needing_data, num_samples=1000, seed=2020)
pdata = ProtocolData(exp_design, data)

In this demo we invoke run_mpi with two extra arguments for robustness across MPI setups and machines with few cores:

• env={'FI_PROVIDER': 'sockets'} can be needed on some MPI distributions. Try without it as well, and omit it unless you need it.

• extra_mpi_args=['--oversubscribe'] lets the launcher start more ranks than the machine has cores. We request num_ranks=3, and many machines (including CI runners) have fewer than three cores; without this flag Open MPI refuses to launch and run_mpi raises a CalledProcessError. Note that --oversubscribe is an Open MPI flag — on MPICH or Intel MPI, omit it (they neither recognize nor require it).

protocol = StandardGST(verbosity=2)
results = protocol.run_mpi(pdata,
    num_ranks=3, mpiexec='auto', env={'FI_PROVIDER': 'sockets'},
    extra_mpi_args=['--oversubscribe']
)

from pygsti.report import construct_standard_report

report = construct_standard_report(
    results, title="MPI Example Report", verbosity=0
)
report.write_html('../../example_files/mpi_example_brief', auto_open=False)