[DO NOT MERGE] Notes os JuMP issues #1218

joaquimg · 2025-01-17T08:17:39Z

main take aways:

1 - use add_to_expression! instead of += (saves allocations of extra expressions)
2 - use add_to_expression!(exp, number, var) instead of add_to_expression!(exp, number * var) because number * var will allocate.
3 - prefer (number * number) * var or var * (number * number) to force computing product of floats first (very very fast) than number * number * var or var * number * number which will depend of the JuMP order of operations (one will be fast, the other slow). use style to keep all coefs in the left and variables in the right. Alsoo use parenthesis when there are multiple coefs.
4 - avoid operations (+, -, /, *, sum) between variables (or varaibles and constants) outside macros, because macros will try to allocate less.
5 - array[:, i] allocates new vectors, use @view to avoid this
6 - attention with type stability when: a) assining values in if-else blocks, b) re-using names inside a function
7 - attention with loop order: julia is column major so inner loops should change the leftmost index.

jd-lara · 2025-01-17T17:15:01Z

src/core/optimization_container.jl

@@ -1627,6 +1627,7 @@ function add_to_objective_invariant_expression!(
 ) where {T <: JuMP.AbstractJuMPScalar}
    T_cf = typeof(container.objective_function.invariant_terms)
    if T_cf <: JuMP.GenericAffExpr && T <: JuMP.GenericQuadExpr
+        # why not add_to...???


before add didn't work for mixed ExpressionTypes. We can re-test

jd-lara · 2025-01-17T17:15:15Z

src/devices_models/devices/common/add_to_expression.jl

@@ -1633,10 +1633,10 @@ function add_to_expression!(
                device_base_power,
            )
            for t in time_steps
-                fuel_expr = variable[name, t] * prop_term_per_unit * dt
                JuMP.add_to_expression!(
                    expression[name, t],


@rodrigomha

for more info:
JuMP.add_to_expression!(exp1, float_coef * jump_var) is worse because it creates an extra temporary expression float_coef * jump_var

jd-lara · 2025-01-17T17:16:21Z

src/devices_models/devices/common/objective_function/market_bid.jl

        JuMP.add_to_expression!(
            gen_cost,
-            cost * multiplier * pwl_var_container[(name, i, time_period)],
+            cost * multiplier,


@rodrigomha

jd-lara · 2025-01-17T17:17:29Z

src/services_models/agc.jl

The AGC functionality is offline for now. We need LHS parameters to revive it

POI is much better nowadays.
Moreover, there is room for improvement, so let me know whenever you want to test stuff and bring it back

jd-lara · 2025-01-17T17:18:11Z

src/services_models/reserve_group.jl

@@ -60,10 +60,12 @@ function add_constraints!(
    for t in time_steps
        resource_expression = JuMP.GenericAffExpr{Float64, JuMP.VariableRef}()
        for reserve_variable in reserve_variables
-            JuMP.add_to_expression!(resource_expression, sum(reserve_variable[:, t]))
+            JuMP.add_to_expression!(resource_expression,


Is this more effective than pre-allocating the expression in 61?

not sure I got what you mean

jd-lara · 2025-01-17T17:19:08Z

src/services_models/reserve_group.jl

        end
        if use_slacks
-            resource_expression += slack_vars[t]
+            JuMP.add_to_expression!(resource_expression, slack_vars[t])


jd-lara · 2025-01-17T17:20:24Z

src/services_models/reserves.jl

@@ -161,19 +161,23 @@ function add_constraints!(
        param = get_parameter_column_refs(param_container, service_name)
        for t in time_steps
            if use_slacks
-                resource_expression = sum(reserve_variable[:, t]) + slack_vars[t]
+                resource_expression = JuMP.@expression(


this could be related to #1000

jd-lara · 2025-01-17T17:20:56Z

src/services_models/reserve_group.jl

These improvements should make it to main

jd-lara · 2025-01-17T17:21:15Z

src/services_models/reserves.jl

These ones are quite important too.

jd-lara

These are great catches @joaquimg I would like to put some of these into main sooner

jd-lara · 2025-01-28T05:16:41Z

src/core/optimization_container.jl

@@ -1627,7 +1627,7 @@ function add_to_objective_invariant_expression!(
 ) where {T <: JuMP.AbstractJuMPScalar}
    T_cf = typeof(container.objective_function.invariant_terms)
    if T_cf <: JuMP.GenericAffExpr && T <: JuMP.GenericQuadExpr
-        # why not add_to...???
+        # TODO: why not add_to...???


The add_to here usually fails is not allowed by JuMP

This is weird, do you have an example? It does not happen in RTS. Maybe it was an old JuMP bug?

This only happens when we have data that mixes PWL and Quadratic cost functions.

If there is any case I can test with those features I'd be happy to.

jd-lara · 2025-01-28T05:17:55Z

src/parameters/update_cost_parameters.jl

    add_to_objective_variant_expression!(container, cost_expr)
+    # TODO: missing _mult below?


no, one expression tracks the cost and the other the fuel consumption

yup not much to do here, as the two need to be kept different

jd-lara · 2025-01-28T05:18:26Z

src/services_models/reserve_group.jl

@@ -61,7 +61,7 @@ function add_constraints!(
        resource_expression = JuMP.GenericAffExpr{Float64, JuMP.VariableRef}()
        for reserve_variable in reserve_variables
            JuMP.add_to_expression!(resource_expression,
-                JuMP.@expression(container.JuMPmodel, sum(reserve_variable[:, t])))
+                JuMP.@expression(container.JuMPmodel, sum(@view reserve_variable[:, t])))


nice trick. @GabrielKS take note

joaquimg · 2025-01-28T06:10:36Z

Feel free to add them in main before I do.
I was trying to catch as many possible issues as I could.

Many of these are general suggestions and should also make it to the JuMP docs.

The gain is not big in the RTS case, about 5% less allocations, but these should make a difference in larger cases with more details.

My plan was to make cleaner PR's after this pass.

joaquimg · 2025-01-28T06:15:19Z

src/devices_models/devices/AC_branches.jl

    for device in devices
        ci_name = PSY.get_name(device)
        if ci_name ∈ PNM.get_radial_branches(radial_network_reduction)
            continue
        end
        limits = get_min_max_limits(device, RateLimitConstraint, U) # depends on constraint type and formulation type
        for t in time_steps
+            # TODO: ternary


ignore this: issue solved recently in MutableArithmetics

joaquimg · 2025-01-28T06:15:30Z

src/devices_models/devices/AC_branches.jl

@@ -606,6 +608,7 @@ function _make_flow_expressions!(
            sum(
                ptdf_col[i] * nodal_balance_expressions[i, t] for
                i in 1:length(ptdf_col)
+                # maybe only sum if nonempy?


joaquimg · 2025-01-28T06:18:45Z

src/devices_models/devices/common/duration_constraints.jl

-                    JuMP.add_to_expression!(lhs_off, (1 - varon[name, i]))
+                    JuMP.add_to_expression!(lhs_off, 1)
+                    JuMP.add_to_expression!(lhs_off, -1, varon[name, i])


the previous allocates more because (1 - varon[name, i]) becomes a new expression

joaquimg · 2025-01-28T06:19:10Z

src/devices_models/devices/common/objective_function/market_bid.jl

@@ -137,11 +137,12 @@ function _get_pwl_cost_expression(
    name = PSY.get_name(component)
    pwl_var_container = get_variable(container, PieceWiseLinearBlockOffer(), T)
    gen_cost = JuMP.AffExpr(0.0)
-    cost_data = PSY.get_y_coords(cost_data)
-    for (i, cost) in enumerate(cost_data)
+    _cost_data = PSY.get_y_coords(cost_data)


this is for type stability

joaquimg · 2025-01-28T06:19:47Z

src/devices_models/devices/electric_loads.jl

-        real = get_variable(container, ActivePowerVariable(), V)[name, t] * pf
-        constraint[name, t] = JuMP.@constraint(jump_model, reactive == real)
+        real = get_variable(container, ActivePowerVariable(), V)[name, t]
+        constraint[name, t] = JuMP.@constraint(jump_model, reactive == real * pf)


the product outside the macro lead to additional allocations

joaquimg · 2025-01-28T06:21:37Z

src/parameters/update_cost_parameters.jl

@@ -131,7 +132,7 @@ function update_variable_cost!(
    end
    mult_ = parameter_multiplier[component_name, time_period]
    variable = get_variable(container, get_variable_type(attributes)(), T)
-    gen_cost = variable[component_name, time_period] * mult_ * cost_data * base_power * dt
+    gen_cost = variable[component_name, time_period] * (mult_ * cost_data * base_power * dt)


this is interesting:
the parenthesis forces the floats the be multiplied before and simplified into a single float. the previous one requires a bunch of operations in an expression. not an allocation issue, but slower than multiplying floats.

joaquimg · 2025-01-28T06:24:44Z

src/services_models/reserves.jl

@@ -224,12 +228,12 @@ function add_constraints!(
            cons[name, t] =
                JuMP.@constraint(
                    jump_model,
-                    var_r[name, t] <= param[t] * requirement * max_participation_factor
+                    var_r[name, t] <= (requirement * max_participation_factor) * param[t]


rule of thumb: first multiply floats and multiply them with jump variables

joaquimg · 2025-01-28T06:25:28Z

src/services_models/reserves.jl

@@ -473,7 +479,7 @@ function add_constraints!(
                PSY.get_active_power_limits(d).min +
                (reserve_response_time - startup_time) * minutes_per_period * ramp_limits.up
        else
-            reserve_limit = 0
+            reserve_limit = 0.0


type stability as the if block would return a float and 0 is ana Int

joaquimg · 2025-01-28T06:41:06Z

src/devices_models/devices/AC_branches.jl

@@ -387,13 +387,15 @@ function add_constraints!(
        slack_lb = get_variable(container, FlowActivePowerSlackLowerBound(), T)
    end

+    # TODO: loop order


this is interesting, and PSI uses mixed patterns.

Julia is column based, so if you vector is array[name, t]
you should iterate as:

for t in T for name in names # do something with array[name, t] end end

which is equivalent to:

for t in T, name in names # do something with array[name, t] end

In the case of sums:

sum(array[name, t] for t in T, name in names)

equivalent to:

sum(array[name, t] for t in T for name in names)

almost equivalent to:

sum(sum(array[name, t] for name in names) for t in T )

so names should typically be the inner loop and t the outer loop.
but if the operation inside the loop is complex both are almost the same.
I would keep the pattern so style would lead you to the correct choices in thigh loops.

Yeah, I originally wrote all the loops in the right order and then had to modify some depending on the context.

joaquimg · 2025-01-28T06:46:16Z

src/devices_models/devices/AC_branches.jl

+        flow_variables_ = @view flow_variables[name, :]
        from_bus = PSY.get_name(PSY.get_from(PSY.get_arc(br)))
        to_bus = PSY.get_name(PSY.get_to(PSY.get_arc(br)))
-        angle_variables_ = ps_angle_variables[name, :]
-        bus_angle_from = bus_angle_variables[from_bus, :]
-        bus_angle_to = bus_angle_variables[to_bus, :]
+        angle_variables_ = @view ps_angle_variables[name, :]
+        bus_angle_from = @view bus_angle_variables[from_bus, :]
+        bus_angle_to = @view bus_angle_variables[to_bus, :]


not sure about these. the @view saves many allocations but the the iteration is worst in the loop below as there are many cache misses.

jd-lara · 2025-01-28T17:45:07Z

Feel free to add them in main before I do. I was trying to catch as many possible issues as I could.

Many of these are general suggestions and should also make it to the JuMP docs.

The gain is not big in the RTS case, about 5% less allocations, but these should make a difference in larger cases with more details.

My plan was to make cleaner PR's after this pass.

I can test this branch on the super large systems and report back w.r.t. to main.

joaquimg · 2025-01-28T19:00:25Z

That would be nice, just build time benchmarking should be enough. no need to run solvers.

[DNM] Notes os JuMP issues

221cdb3

joaquimg mentioned this pull request Jan 17, 2025

Remove the warning in operator_warn jump-dev/JuMP.jl#3910

Closed

jd-lara reviewed Jan 17, 2025

View reviewed changes

src/services_models/reserve_group.jl

Copy link

Member

jd-lara Jan 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These improvements should make it to main

jd-lara reviewed Jan 17, 2025

View reviewed changes

src/services_models/reserves.jl

Copy link

Member

jd-lara Jan 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These ones are quite important too.

jd-lara reviewed Jan 17, 2025

View reviewed changes

more notes

363e07e

jd-lara reviewed Jan 28, 2025

View reviewed changes

joaquimg commented Jan 28, 2025

View reviewed changes

joaquimg mentioned this pull request Jan 28, 2025

Suggestions for documentation improvements jump-dev/JuMP.jl#2348

Open

13 tasks

odow mentioned this pull request Jan 30, 2025

Fix mutable allocations in n-ary multiplication jump-dev/MutableArithmetics.jl#326

Merged

		add_to_objective_variant_expression!(container, cost_expr)
		# TODO: missing _mult below?

[DO NOT MERGE] Notes os JuMP issues #1218

Are you sure you want to change the base?

[DO NOT MERGE] Notes os JuMP issues #1218

Conversation

joaquimg commented Jan 17, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jd-lara Jan 17, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jd-lara left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

joaquimg commented Jan 28, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jd-lara commented Jan 28, 2025

joaquimg commented Jan 28, 2025

joaquimg commented Jan 17, 2025 •

edited

Loading

jd-lara Jan 17, 2025 •

edited

Loading

joaquimg commented Jan 28, 2025 •

edited

Loading