I first became aware of processor group awareness when working in Sketchup with Vray. They have since dealt with the issue properly, but for a time being you either had to work in a hack in the form of adding an environmental variable, else disable some cores and NUMA. This is now the case when I’m working with Unreal 4.
So I won’t get into the technicals, as I never cared to fully understand them anyways. What I learned is that issues often arise when working in programs on rigs that have over 64 logical cores. That means anything over a dual-cpu setup with 16 physical cores and hyperthreading. Apparently a limitation within Microsoft gives you up to this value in cores before splitting you into a second processor group.
My case has 72 and 80 core rigs. What behavior I witnessed was lightmass happening on either 1 processor group or the other. So depending on if NUMA was enabled or not, I would get for a 72 core system 64 or 8 cores working on lightmass, else 50% of the cores. There appeared to be no reason to which rig would get however many cores with lightmass, until I realized the 80 core rig would split 64 or 16, not 64 or 8. That lead me back to the PGA issue in the past, and I disabled both NUMA and cores until I was with 1 processor group of 64 cores on each rig.
Now each lightmass swarm rendering will fully utilize the same core count on each rig. This appears to be the best that is available for the moment. I couldn’t find anything else in search releating to this, so decided to make a thread in hopes it’ll save someone time in the future. If you have more than 64 logical cores on your rigs, you will need to get your rig down to 1 processor group. You can disable HT, but I would recommend disabling cores over that unless you’re dealing with a dual socket 2679V4 (24 physical cores).