Affine expressions are slow

[odow]

@jd-lara has a problem where creating affine expressions is a bottleneck.

There are two main causes.

This double-hash in _add_or_set!

JuMP.jl/src/aff_expr.jl

Lines 18 to 28 in 2ff880d

	# Utilities for OrderedDict
	function _add_or_set!(dict::OrderedDict{K,V}, k::K, v::V) where {K,V}
	# Adding zero terms to this dictionary leads to unacceptable performance
	# degradations. See, e.g., https://github.com/jump-dev/JuMP.jl/issues/1946.
	if iszero(v)
	return dict # No-op.
	end
	# TODO: This unnecessarily requires two lookups for k.
	dict[k] = get!(dict, k, zero(V)) + v
	return dict
	end

This size hint:

JuMP.jl/src/aff_expr.jl

Lines 390 to 401 in 2ff880d

	function add_to_expression!(
	aff::GenericAffExpr{C,V},
	coef::_Constant,
	other::GenericAffExpr{C,V},
	) where {C,V}
	sizehint!(aff, length(linear_terms(aff)) + length(linear_terms(other)))
	for (term_coef, var) in linear_terms(other)
	_add_or_set!(aff.terms, var, coef * term_coef)
	end
	aff.constant += coef * other.constant
	return aff
	end

In particular, the size hint is pessimistic in that it suggests reserving enough space for the union of the elements assuming that there are no duplicates.

[odow]

Closing because this is not a JuMP problem.

The real issue is that there is something to do with the affine expressions inside ParameterJuMP that are causing a large slow-down: JuliaStochOpt/ParameterJuMP.jl#85