LinearExpression cache doesn't do anything?

[michaelbynum]

Summary

The cache for LinearExpression properties linear_vars and linear_coefs is a class attribute, which means that it always gets overwritten, making those properties quite slow unless you only have one linear expression. Here is an example:

import pyomo.environ as pe
from pyomo.common.timing import TicTocTimer
import sys


def get_vars(a, b):
    res = list(a.linear_vars)
    res.extend(b.linear_vars)
    return res


def main():
    m = pe.ConcreteModel()
    m.a = pe.Set(initialize=list(range(10)))
    m.x = pe.Var(m.a)
    m.y = pe.Var(m.a)
    e1 = sum(i*v for i, v in enumerate(m.x.values()))
    e2 = sum(i*v for i, v in enumerate(m.y.values()))

    e1.linear_vars
    e2.linear_vars

    n = int(sys.argv[1])

    timer = TicTocTimer()
    timer.tic('start')
    for i in range(n):
        get_vars(e1, e1)
    timer.toc('done with a single expression')

    for i in range(n):
        get_vars(e1, e2)
    timer.toc('done with two expressions')


if __name__ == '__main__':
    main()

Output with n=1,000,000:

[    0.00] start
[+   0.30] done with a single expression
[+   2.37] done with two expressions

I think that we should deprecate those properties and encourage people to use the args property.