bug#67502: [Cuirass] ‘cuirass remote-worker’ gets the CPU count wrong on the OverDrive

bug#67502: [Cuirass] ‘cuirass remote-worker’ gets the CPU count wrong on the OverDrive

[Ludovic Courtès]

On the OverDrive (AArch64), ‘cuirass remote-worker’ (1.2.0-1.bdc1f9f) says:

  starting 2 workers (parallelism: 1 cores) for server at 10.0.0.1

Instead it should use two cores for each worker:

--8<---------------cut here---------------start------------->8---
ludo@dover ~$ guile -c '(use-modules (ice-9 threads)) (pk (current-processor-count))'

;;; (4)
ludo@dover ~$ guile -c '(use-modules (ice-9 threads)) (pk (ceiling-quotient (current-processor-count) 2))'

;;; (2)
ludo@dover ~$ nproc
4
--8<---------------cut here---------------end--------------->8---

Since ‘current-processor-count’ is implemented indirectly in terms of
‘sched_getaffinity’, this suggests that the process starts with a bogus
affinity mask.  (Time passes…)  That’s indeed the case:

--8<---------------cut here---------------start------------->8---
ludo@dover ~$ sudo herd status cuirass-remote-worker
Status of cuirass-remote-worker:
  It is started.
  Running value is 21279.
  It is enabled.
  Provides (cuirass-remote-worker).
  Requires (avahi-daemon guix-daemon networking).
  Will be respawned.
ludo@dover ~$ guile -c '(pk (getaffinity 21279))'

;;; (#*1000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000)
--8<---------------cut here---------------end--------------->8---

Compare to the affinity mask on x86_64-linux-gnu:

--8<---------------cut here---------------start------------->8---
ludo@guix-hpc3 ~$ sudo guile -c '(pk (getaffinity 1817))'

;;; (#*1111111111111111111111111111111100000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000)
--8<---------------cut here---------------end--------------->8---

Interesting that the initial affinity mask differs on aarch64-linux-gnu
compared to x86_64-linux-gnu.

Ludo’.

bug#67502: [Cuirass] ‘cuirass remote-worker’ gets the CPU count wrong on the OverDrive

[Ludovic Courtès]

Ludovic Courtès <ludovic.courtes@inria.fr> skribis:

> ludo@dover ~$ sudo herd status cuirass-remote-worker
> Status of cuirass-remote-worker:
>   It is started.
>   Running value is 21279.
>   It is enabled.
>   Provides (cuirass-remote-worker).
>   Requires (avahi-daemon guix-daemon networking).
>   Will be respawned.
> ludo@dover ~$ guile -c '(pk (getaffinity 21279))'
>
> ;;; (#*1000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000)

This was due to ‘run-fibers’ binding one thread per CPU core.  Thus,
calling ‘getaffinity’ from within ‘run-fibers’ shows only one CPU and
likewise ‘current-processor-count’ returns 1.

Fixed in Cuirass commit 87a6d6ea7ae79fdf487bbcfd44bb3dce2d7c6e82.

Ludo’.