Windows CP-SAT: Significant Drop in CPU Utilization After Initial Solutions

[gregy4]

What version of OR-Tools and what language are you using?
Version: v99bugfix (9.13.4637)
Language: C++, Java

Which solver are you using (e.g. CP-SAT, Routing Solver, GLOP, BOP, Gurobi)
CP-SAT

What operating system (Linux, Windows, ...) and version?
Windows 10, 11

What did you do?
Steps to reproduce the behavior:
Execute attached proto file with 32 search workers.

What did you expect to see
More or less consistent cpu usage.

What did you see instead?
After first two solutions are found, there's an interval over one minute (4 minutes on threadripper 1950x) where cpu usage is significantly lower. Ryzen master shows significantly lower consumption and usage of less than 8 cores, task manager shows similar values in percentage of used cpu. I was able to reproduce same behaviour back to 9.11. In version 9.10 it was not visible.

With 16 search workers it was not visible to reproduce it, with 24 workers the interval was much shorter (under 10 seconds) and I don't see it as a problem here.

It was not possible to reproduce it in linux (debian 12).

Make sure you include information that can help us debug (full error message, model Proto).

model.zip

Anything else we should know about your project / environment

[lperron]

thanks. Hard to reproduce, but we will look at it

[lperron]

On linux, 64 workers. I see a 5 seconds drop.
My guess is memory unallocation from first solution workers. Windows is notoriously slow in that scenario.

We will check if we can pinpoint the data structure that is slow to delete.

Thanks

[gregy4]

So far I found a bug in deallocation of memory. According to https://learn.microsoft.com/en-us/cpp/c-runtime-library/reference/aligned-malloc?view=msvc-170 should be memory freed by _aligned_free for _aligned_malloc used in util/aligned_memory_internal.h for Windows case.

Unfortunately this didn't helped me to solve the issue.

[gregy4]

Lower cpu utilization started with commit ffbc8d1. With some effort I was able to backport fixes needed for windows compilation. Before the commit was also visible lower cpu utilization but not so low.

[lperron]

Yes, helper classes for violation ls use a lot of small protos and small allocations. We just improved it for no_overlap_2d. We need to do it for no_overlap. Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le mer. 2 avr. 2025 à 23:44, gregy4 ***@***.***> a écrit :

…

Lower cpu utilization started with commit ffbc8d1 <ffbc8d1>. With some effort I was able to backport fixes needed for windows compilation. Before the commit was also visible lower cpu utilization but not so low. — Reply to this email directly, view it on GitHub <#4615 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3I57YTFYN6VIXPL4UT2XRK4XAVCNFSM6AAAAAB2FHZZO2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDONZTHAYDOMJWHE> . You are receiving this because you were assigned.Message ID: ***@***.***> [image: gregy4]*gregy4* left a comment (google/or-tools#4615) <#4615 (comment)> Lower cpu utilization started with commit ffbc8d1 <ffbc8d1>. With some effort I was able to backport fixes needed for windows compilation. Before the commit was also visible lower cpu utilization but not so low. — Reply to this email directly, view it on GitHub <#4615 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3I57YTFYN6VIXPL4UT2XRK4XAVCNFSM6AAAAAB2FHZZO2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDONZTHAYDOMJWHE> . You are receiving this because you were assigned.Message ID: ***@***.***>

[lperron]

Can you check if the situation is improved ?

[gregy4]

Unfortunately in Windows situation didn't changed. In Linux I see some improvement during 5 seconds drop.