DragonFlyBSD Gets Performance Tuning For Threadripper 2990WX Topology, Scheduler Tuning

While it was just days ago that DragonFlyBSD lead developer Matthew Dillon got his hands on a Threadripper 2990WX 32-core / 64-thread "beast", got it working under this long ago forked operating system from FreeBSD, and proceeded to exclaim with joy how powerful this system is, he's now made it even better. Dillon has landed some additional kernel work to benefit the AMD Ryzen Threadripper 2990WX.

On top of the kernel changes made at the end of last week to bring-up the 2990WX support, Matthew Dillon has now had the time to do some tuning to make this 64-thread system perform even faster. Hitting DragonFlyBSD Git overnight was a patch to update the AMD topology detection as seeing four nodes with eight cores and two threads per core, per node. Previously it was just exposed as a CPU with 32 cores and 2 threads per core.

The DragonFly patch also now is able to expose how much memory is accessible from each node, an instability fix in the kernel's scheduler when dealing with large core counts, and memory-on-node weighting in the scheduler.

These changes should help the Threadripper 2990WX performance even more particularly for multi-threaded workloads with the scheduler cha nge and the memory-on-node weighting will give initial scheduling preference to nodes with more RAM attached to them and begin scheduling to the nodes with no direct memory attached as the load increases.

It was just yesterday I wrapped up some DragonFlyBSD benchmarks on the 2990WX for an imminent article but looks like will be doing some extra DragonFly tests in the days ahead. All the technical details in this commit.