Linux: detectCores(logical = FALSE) is always equal to detectCores(logical = TRUE)

[HenrikBengtsson]

Issue

On Linux systems(*), parallel::detectCores() ignores the logical argument. For example, on my notebook with four physical CPU cores with two "processors" each, I get:

> parallel::detectCores(logical = FALSE)
[1] 8
> parallel::detectCores(logical = TRUE)  ## default
[1] 8

I think the former should really be 4.

FWIW, help("detectCores", package = "parallel") acknowledges this:

logical | Logical: if possible, use the number of physical CPUs/cores (if FALSE) or logical CPUs (if TRUE). Currently this is honoured only on macOS, Solaris and Windows.

(*) A Linux system is where R.version$os starts with linux, e.g. in R 4.2.2 on my Ubuntu 20.04 system I have R.version$os == "linux-gnu".

Troubleshooting

In both cases, detectCores() uses the following system(..., intern = TRUE) call on Linux:

$ grep "^processor" /proc/cpuinfo 2>/dev/null | wc -l
8

to get the number of cores - logical or not.

If I look at lscpu (not installed on all systems), I get:

$ lscpu 
…
CPU(s):                          8
On-line CPU(s) list:             0-7
Thread(s) per core:              2
Core(s) per socket:              4
Socket(s):                       1
…

on my machine. That suggests there are 8 logical CPU cores, and 4*1 = 4 physical ones. We can query the raw data for this from /proc/cpuinfo as:

$ cat /proc/cpuinfo | grep -E "^(processor|core id)"
processor	: 0
core id		: 0
processor	: 1
core id		: 1
processor	: 2
core id		: 2
processor	: 3
core id		: 3
processor	: 4
core id		: 0
processor	: 5
core id		: 1
processor	: 6
core id		: 2
processor	: 7
core id		: 3

which explains:

$ grep "^processor" /proc/cpuinfo 2>/dev/null | wc -l
8

for counting the number of logical CPU cores. However, for the physical ones, I think we should count the unique number of core IDs;

$ grep "^core id" /proc/cpuinfo 2>/dev/null | sort -u | wc -l
4

Action

So, this begs the question, why doesn't parallel::detectCores() do this? I'm pretty sure this has been discussed somewhere before. The first task is to identify any discussions and rationales for the current implementation.

See also

• This is related to Issue detectThreadsPerCore(): infer how many threads per cores the machine's CPU provides #47 (detectThreadsPerCore())
• It's complicated, e.g. https://superuser.com/questions/378273/strange-cpuinfo-how-many-cores-does-these-physical-cpu-has/932418#932418
• R-devel thread 'Detecting physical CPUs in detectCores() on Linux platforms' started on 2023-08-07 (https://stat.ethz.ch/pipermail/r-devel/2023-August/082772.html).