On 3/29/22 09:46, Jagdish Gediya wrote: >> I'd love to see some more information about *why* the current code >> doesn't work. Is it purely a bug or was it mis-designed? > I think the design seems to be intended because as per the comment > in the current code, it was known that there are limits to the node > sharing as a demotion target. Ahh, that makes sense. You've got some slow memory that's not really associated closely enough with an initiator node to be used exclusively for demotions from that node. The original code was really intended for demotions when the system is made up of relatively identical building blocks, like one CPU with two kinds of memory attached. I think what you're doing here makes sense. The only thing I'll say is that, at some point, we need to stop messing with the demotion order that the kernel comes up with. It's not going to be able to handle *everything* perfectly. We're probably not at the point that we should stop enhancing the kernel-generated demotion order, but let's at least keep in mind that we should stop eventually.
