search

jump-dev / HiGHS.jl

A Julia interface to the HiGHS solver
https://highs.dev
MIT License
97 stars 13 forks source link

Set Threads on Windows by default #161

Closed odow closed 1 year ago

odow commented 1 year ago

Closes #160

Users can still over-ride this manually.

codecov[bot] commented 1 year ago

Codecov Report

Patch coverage has no change and project coverage change: -0.06 :warning:

Comparison is base (bbf2835) 80.24% compared to head (b9096d5) 80.19%.

Additional details and impacted files ```diff @@ Coverage Diff @@ ## master #161 +/- ## ========================================== - Coverage 80.24% 80.19% -0.06% ========================================== Files 3 3 Lines 1554 1555 +1 ========================================== Hits 1247 1247 - Misses 307 308 +1 ``` | [Impacted Files](https://codecov.io/gh/jump-dev/HiGHS.jl/pull/161?src=pr&el=tree&utm_medium=referral&utm_source=github&utm_content=comment&utm_campaign=pr+comments&utm_term=jump-dev) | Coverage Δ | | |---|---|---| | [src/MOI\_wrapper.jl](https://codecov.io/gh/jump-dev/HiGHS.jl/pull/161?src=pr&el=tree&utm_medium=referral&utm_source=github&utm_content=comment&utm_campaign=pr+comments&utm_term=jump-dev#diff-c3JjL01PSV93cmFwcGVyLmps) | `91.35% <0.00%> (-0.08%)` | :arrow_down: | Help us with your feedback. Take ten seconds to tell us [how you rate us](https://about.codecov.io/nps?utm_medium=referral&utm_source=github&utm_content=comment&utm_campaign=pr+comments&utm_term=jump-dev). Have a feature suggestion? [Share it here.](https://app.codecov.io/gh/feedback/?utm_medium=referral&utm_source=github&utm_content=comment&utm_campaign=pr+comments&utm_term=jump-dev)

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Do you have feedback about the report comment? Let us know in this issue.

jajhall commented 1 year ago

Users can still over-ride this manually.

They can and, for now, they won't notice any performance degradation.

However, this won't be the case once we've got the concurrent LP solver, parallel MIP search and parallel IPM solver.

odow commented 1 year ago

They can and, for now, they won't notice any performance degradation. However, this won't be the case once we've got the concurrent LP solver, parallel MIP search and parallel IPM solver.

So just to summarize my understanding (let me know if I have some part wrong):

The stall on Windows when the default is 0 but not 1 seems like the bug is in HiGHS, not in how we're compiling things on Yggdrasil. But it's pretty hard to track down because we can't compile Yggdrasil binaries directly on windows machines, and I don't have a Windows machine to test things on.

jajhall commented 1 year ago

So just to summarize my understanding (let me know if I have some part wrong):

  • The default is 0

Yes

  • When the value is 0, HiGHS will internally set threads to half the number of cores

Half the maximum number of threads for the machine

  • When the value is 0, HiGHS sometimes stalls on Windows only

Yes

  • If the user manually sets a positive value, no stall happens

Not sure: I can try some experiments on my W10 laptop

  • Currently HiGHS doesn't really use threading, so setting to 1 should have minimal impact

Yes

  • It will have an impact once HiGHS starts to really use multithreading.

Yes

The stall on Windows when the default is 0 but not 1 seems like the bug is in HiGHS, not in how we're compiling things on Yggdrasil. But it's pretty hard to track down because we can't compile Yggdrasil binaries directly on windows machines, and I don't have a Windows machine to test things on.

It would seem to be something to do with how HiGHS shuts down threads on Windows. It would be good to identify why, but there's too much to do at the moment. Using 1 is certainly OK with JuMP. HiGHS will use a larger default value so that our Linux benchmark performance for the MIP solver isn't compromised.

odow commented 1 year ago

Would doing something like:

julia> Sys.cpu_info()
8-element Vector{Base.Sys.CPUinfo}:
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz    3248726 s          0 s    1526146 s   22719837 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz     226257 s          0 s     241621 s   27026610 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz    2975710 s          0 s    1113774 s   23405012 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz     219407 s          0 s     224294 s   27050788 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz    2609224 s          0 s     887438 s   23997833 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz     220094 s          0 s     209349 s   27065045 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz    2662113 s          0 s     782994 s   24049388 s          0 s
 Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz: 
        speed         user         nice          sys         idle          irq
     2300 MHz     227937 s          0 s     199417 s   27067132 s          0 s

julia> div(length(Sys.cpu_info()), 2)
4

be a better default?

jajhall commented 1 year ago

So long as 4 doesn't mess up Windows, it's the value that I was thinking of for the in-HiGHS default.

odow commented 1 year ago

So long as 4 doesn't mess up Windows, it's the value that I was thinking of for the in-HiGHS default.

Let's put this out there and see if any complaints roll in 😆. That's probably the only way we'll know.