Forum Discussion
How to reduce memory usage by the solver?
We have a large model and it ran out of memory when it is starting concurrent solve with dual and barrier method. So is it possible to just pick one method here to reduce the memory requirement? Thank you.
=============== log output ============================
FICO Xpress v9.3.0, Hyper, solve started 11:51:09, Apr 17, 2024
Heap usage: 56GB (peak 56GB, 10GB system)
Maximizing MILP PCTO_Duration using up to 32 threads and up to 251GB memory, with these control settings:
OUTPUTLOG = 1
Original problem has:
76783959 rows 111750376 cols 548702657 elements 111750376 entities
Presolved problem has:
8539375 rows 37089513 cols 136913748 elements 37089513 entities
LP relaxation tightened
Presolve finished in 1729 seconds
Heap usage: 71GB (peak 126GB, 10GB system)
Coefficient range original solved
Coefficients [min,max] : [ 2.62e-05, 5.78e+02] / [ 1.55e-12, 1.68e+00]
RHS and bounds [min,max] : [ 1.00e+00, 1.00e+00] / [ 1.00e+00, 4.41e+03]
Objective [min,max] : [ 3.90e-01, 3.09e+04] / [ 2.02e-01, 5.39e+04]
Autoscaling applied standard scaling
Symmetric problem: generators: 55, support set: 168
Number of orbits: 58, largest orbit: 18
Row orbits: 29, row support: 84
Will try to keep branch and bound tree memory usage below 176.1GB
*** Solution found: 366408331.8 Time: 1946.44 Heuristic: T ***
*** Solution found: 1539602960. Time: 1982.51 Heuristic: e ***
*** Solution found: 1539617893. Time: 2013.03 Heuristic: h ***
*** Solution found: 1539630086. Time: 2046.97 Heuristic: i ***
Starting concurrent solve with dual (1 thread) and barrier (31 threads)
?20 Error: Out of memory (0)
?20 Error: Out of memory (0)
?20 Error: Out of memory (0)
5 Replies
You can use the -d flag to use only dual simplex or the -b flag to use only the barrier algorithm for the root node.
On a related note, it seems like the presolver managed to remove a large portion of the problem. This suggests that the model could probably be improved, so the initial model isn't as large.
Thank you for the reply. We are using Python to formulate and solve the model. Can you provide the Python code for setting this flag? Thank you!
You can use the LPFLAGS control:
https://www.fico.com/fico-xpress-optimization/docs/latest/solver/optimizer/HTML/LPFLAGS.html
You can set it to 2 for barrier or 0 for dual simplex. You can use the setControl method on an xpress problem to set the control value:
https://www.fico.com/fico-xpress-optimization/docs/latest/solver/optimizer/python/HTML/problem.setControl.html
It seems there is a little confusion about how the LPFLAGS bit control works: This control allows for a custom mix of root LP algorithms.
You activate dual simplex by setting bit 0 (value 1) to 1, you activate barrier by setting bit 2 (value 4) to 1. Therefore, setting LPFLAGS = 4 activates barrier, but disables dual simplex. LPFLAGS = 1 activates only dual simplex. Mixing both (the current behavior) is achieved by setting LPFLAGS = 5.
In python, you therefore use
p = xpress.problem() p.setControl('lpflags', 1) # or 4to only run your favorite LP algorithm.
It is common that one algorithm solves a root LP faster than the other, so it is a good practice to choose that algorithm only.
Thank you very much! This is what I need.
