https://bugs.winehq.org/show_bug.cgi?id=43277
--- Comment #32 from Ivan Kalvachev iive@yahoo.com --- Any new development on this issue?
Few days ago we talked on IRC and we did a bit of experimenting.
I think I found mistake in one of the experiments we did.
It seems that the NUMA nodes, or as AMD calls them CCX are formed by 4 physical cores. (I thought they are 2 previously).
On your processor Ryzen 1600, you have 2x CCX and each one has 3 physical cores (one is disabled).
So the layout looks like this: ccx/node0 { (0,1),(2,3),(4,5) } ccx/node1 { (6,7),(8,9),(10,11) }
Here, (0,1) are cpu's threads that run on same physical core.
So, when you disabled half of the cores, you disabled one whole ccx/node and this avoids the slow interconnect between nodes.
When I asked you to set affinity for wineserver and fc3 on cpu 5 & 7, I thought they would be on the same node, but they are actually on different ones.
Could you repeat the experiment, but this time with cpu3&5 ? Or just set affinity to the first 6 cores on all wine related programs (wineserver and everything that is *.exe).