Comment by malkia

3 years ago

It was all good, until NUMA came, and now you have to careful rethought your process, or you get lots of performance issues in your (otherwise) well threaded code. Speaking from first-hand experience, when our level editor ended up being used by artists on a server class machine, and supposedly 4x faster machine was actually going 2x slower (why, lots of std::shared_ptr<> use on our side, or any atomic reference counting) caused slowdowns, as the cache (my understanding) had to be synchronized between the two physical CPUs each having 12 threads.

But really not the only issue, just pointing out - that you can't expect everything to scale smoothly there, unless well thought, like ask your OS to allocate your threads/memory only on one of the physical CPUS (and their threads), and somehow big disconnected part of your process(es) on the other one(s), and make sure the communication between them is minimal.. which actually wants micro-services design again at that level.