Re: [PATCH 21/28] generic/531: limit max files per CPU

["Nirjhar Roy (IBM)" <nirjhar.roy.lists@gmail.com>]

On Thu, 2025-04-17 at 13:01 +1000, Dave Chinner wrote:
> From: Dave Chinner <dchinner@redhat.com>
> 
> Currently g/531 runs t_open_files on every CPU, and with default
> kernel settings that means 50,000 files per CPU are tested. On 64p
> machines this means the test tries to create and unlink over 3
> million files. This takes a long time:
> 
> Ten slowest tests - runtime in seconds:
> generic/531 534
> .....
> 
> Yet generic/531 is included in the 'quick' test group. It is
> anything but "quick" on large CPU count systems.
> 
> Further, small filesystems  like are typically used for fstests do
> not have the inherent concurrency to scale out this workload
> effectively. Even using the mkfs.xfs concurrency options requires
> using >250GB scratch devices on 64p machines because it won't make
> AGs smaller than 4GB. Hence to get 64-way concurrency in the
> filesystem, we need huge devices to be set up, and that's not really
> practical for check-parallel.
> 
> Hence limit the total number of files this test will create
> to a sane number, and distribute them over all the CPUs so that
> the test runtime does not blow out on big systems. LOAD_FACTOR can
> still be used to increase runtime of the test by increasing the
> total number of files created.
> 
> Limiting the total number of files created brings g/531
> system back into the "quick" test range on a 64p system:
> 
> generic/531        5s
Right, on my system with 16p, I too noticed the reduced test time. 
generic/531 24s ...  199s (24s with this change, 199s on for-next)
Reviewed-by: Nirjhar Roy (IBM) <nirjhar.roy.lists@gmail.com>
--NR
> 
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> ---
>  tests/generic/531 | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/tests/generic/531 b/tests/generic/531
> index 07dffd9fd..3f691c0f8 100755
> --- a/tests/generic/531
> +++ b/tests/generic/531
> @@ -29,13 +29,13 @@ _scratch_mount
>  # Try to load up all the CPUs, two threads per CPU.
>  nr_cpus=$(( $(getconf _NPROCESSORS_ONLN) * 2 ))
>  
> -# Set ULIMIT_NOFILE to min(file-max / $nr_cpus / 2, 50000 files per LOAD_FACTOR)
> +# Set ULIMIT_NOFILE to min(file-max / 2, 100000) / $nr_cpus files per LOAD_FACTOR)
>  # so that this test doesn't take forever or OOM the box
> -max_files=$((50000 * LOAD_FACTOR))
> -max_allowable_files=$(( $(cat /proc/sys/fs/file-max) / $nr_cpus / 2 ))
> +max_files=$((100000 * LOAD_FACTOR))
> +max_allowable_files=$(( $(cat /proc/sys/fs/file-max) / 2 ))
>  test $max_allowable_files -gt 0 && test $max_files -gt $max_allowable_files && \
>  	max_files=$max_allowable_files
> -ulimit -n $max_files
> +ulimit -n $((max_files / nr_cpus))
>  
>  # Open a lot of unlinked files
>  echo create >> $seqres.full