On 6/18/26 12:16, vermaden wrote:
>> amd64 system with 16 2-thread cores (SMT):
>>
>> # echo $(nproc)
>> 32
>>
>> (No cpuset based subsetting involved.)
>>
>> # sysctl kern.smp.cores
>> kern.smp.cores: 16
> 
> I was probably not specific enough.
> 
> I would set GELI to use number of threads equal to physical cores - no matter 
> where the info about physical cores comes from ... and that is for ALL 
> encrypted GELI devices - not 'per device'.

nproc does not report physical cores in amd64 SMT/Hyper-Threading
contexts unless cpuset has been used to produced 1 thread per core.
kern.smp.cores does return physical cores.


Also, thinking about the default and non-amd64 contexts:

POWER8 and later families have more like 4 or 8 threads per core, that
do not have the same tradeoffs in that design. A justification based on
am64 SMT/Hyper-Threading behavior being biased to single thread
performance likely would not apply to the likes POWER8 or later as a
justification, for example.

I wonder if the default for such tradeoffs likely should be
architecture/platform specific?

> 
> So If you have ONE disk and 4 physical cores - it means use 4 GELI threads.
> 
> If You have 80 disks and 10 physical cores - it means use 10 GELI threads for 
> ALL of these encrypted disks - NOT 10 threads per disk.
> 
> Hope that helps.
> 
> Regards,
> vermaden
> 
> 


-- 
===
Mark Millard
marklmi at yahoo.com

Reply via email to