Performance of Parallel Tempering

[joelgdwave]

Compared to the BQM solver in HSS, ParallelTempering is very slow. I've attached a Python program for observing this, including a DQM from Accenture.

import hybrid
from dimod import BinaryQuadraticModel
bqm = ... 
main = hybrid.reference.HybridizedParallelTempering(num_replicas=10, max_iter=None, max_time=None, convergence=3, num_sweeps=10000)
init_state = hybrid.State.from_sample(hybrid.utils.min_sample(bqm), bqm)
solution = main.run(init_state).result().samples.first
print(solution)

I have observed output like this:

real	1m16.454s
user	1m25.997s
sys	0m2.610s

whereas BQM solver in HSS requires times like this:

real	0m11.247s
user	0m1.875s
sys	0m0.573s

I will send the BQM in a file to @randomir; I don't know yet how to clip a very long string into Github Markdown.