Re: Solving thundering Herd with glibc...

Am 26.04.2025 um 23:41 schrieb Chris M. Thomasson:

On 4/26/2025 12:25 AM, Bonita Montero wrote:

>
I just tried to awaken all 31 threads from outside holding the mutex,
but not from inside:
>
     for( size_t r = N; r; --r )
     {
         unique_lock lock( mtx );
         signalled = nClients;
         ai.store( nClients, memory_order_relaxed );
         lock.unlock();
         if( argc < 2 )
             cv.notify_all();
         else
             for( int c = nClients; c; cv.notify_one(), --c );
         bs.acquire();
     }
>
The result: 7.500 context switches per thread, not 3.000.
>
     10000 rounds,
     7498.06 context switches pe thread
>
So never signal a condvar to multiple threads from outside !

So, do that. It's your software. Do what you like. This is a very old debate. Take your contrived test and just, roll with it. Whatever.

There's nothing to debate, I've measured it: If you awake a single
thread it doesn't matter if you awake from inside or outside, if you
awake multiple threads awakening them from inside is multiple times
faster.