On 25/06/2020 18:37, Adam Borowski wrote:
I'm afraid that `nproc` shows only the number of _currently_ online CPUs,
which on mobile processors tends to be 1 when starting a job.  As there's
a need to conserve power, holding cores online when they have nothing to
do would be a waste, thus they constantly get onlined and offlined.

For example, on a 10-core box where the processor consists of three
clusters, 4+4+2 cores, topline graph shows:

(oooo▁▁▁ oo)
(oooo▁ ▁ oo)
(oooo▁▁▁ oo)
(oooo▁ ▁ oo)
(oooo▁▁  oo)
(oooo▁  ▁oo)
(oooo▁▁  oo)
(oooo▁   oo)

(1 line = 1 second, "o" means offline core, otherwise it's utilization

Because of noisy GUI (that eg. draws this very graph), the machine keeps
flipping between onlining just CPU 0, or a cluster of CPU4..7.  Thus,
`nproc` will randomly says either "1" or "4", while the user expects to
run her compile with all 10 cores.

I found out that numa_num_task_cpus() from libnuma gets the right answer,
handling both affinity mask and CPUs present-but-currently-offline.  But
alas, it doesn't provide a command-line tool, and thousands of scripts
already use nproc, thus switching a tool would be a waste of effort.

Ie, could you please make nproc include all available CPUs rather than
only online ones?

Does `nproc --all` suffice for your use case?

Reply via email to