The machine that https://friendica.eskimo.com runs on is an i9-10980xe, this is an 18 core, 36 thread processor, all cores overclocked at 4.5Ghz with 256G of RAM.

Typically the CPU is about 10% occupied and 90% idle, but still sometimes processes are slow or stall. Disk I/O likewise is not saturated, and out of 256G of RAM typically about half is in play and half is available.

So I am doing an experiment in a bit in which I am using NUMA emulation to divide the memory up into four fake NUMA nodes and treating the memory as if a NUMA machine even though it isn't. The idea is that by having four parallel memory controllers instead of one, memory allocation lock contention will be greatly reduced, and since there is no cross-node penalty I hope this won't be expensive and will pay off in better performance.

We will find out in about an hour.