On Thu, May 15, 2025 at 02:06:39PM -0400, Ted Unangst wrote:
> On 2025-05-14, Martin Pieuchot wrote:
> > Instead of adding dead threads at the head of the deadproc list insert
> > them at the tail.
> > 
> > This improve latency and reduce variance when building with 24 CPUs.
> 
> Any theories why? I would think, if it mattered, a LIFO would be more
> cache friendly. 

Because signaling to the parent is done in the reaper right now. So if you
have sh executing printf and then 20 cc exit right after you want printf
to be reaped first and not wait for 20 cc processes to cleanup.
I think overall a fifo is fairer in that regard since the first entry can
not starve on the queue. 

-- 
:wq Claudio