News for hackers. Stuff that matters.
Re:Microsoft blames monster email outage on software defect ... by malkia on Tuesday November 17, 2026 @12:08PM (#50534559)
Lock contention, especially cache line contention in large NUMA systems. If hardware was upgraded to NUMA machines, or VM's were not cordoned each to sit in each it's own NUMA node (CPU, Caches & Memory) then something like this would happen.

We saw with our own tools back when an artist were given a low-core "server" like machine with many CPUs (good we thought), and 2 NUMA nodes, and suddely the level editor for our game was running much slower. We could blame std::shared_ptr<> all day long, or similar but it's there and it was real.

Tried the 4kb interleaved paging, but then everything was equally slow. At the end we simply blocked the App to run only on one NUMA node (still using memory from the other one), and things were back to "normal". "Normal", because promises were not kept - we could not use all CPUs.

There are no comments on this comment yet.