I think the interesting part here is that it's not just some 'compressed pages' taking up an arbitrary fraction of RAM (regardless of how much more efficient their reimplementation of zswap is), but being dynamically tuned based on machine learning (Gaussian processes) to balance savings with performance objectives in the real world. If you have that framework, I could see this being extended to forms of memory other than local on-mobo RAM, like Flash drives, or actual RAM elsewhere in the datacenter: they just have worse performance characteristics than local RAM and get used less, is all.
Why isn't this just called "compressed pages"? Software-defined far memory isn't really needed, is it?
I think the interesting part here is that it's not just some 'compressed pages' taking up an arbitrary fraction of RAM (regardless of how much more efficient their reimplementation of zswap is), but being dynamically tuned based on machine learning (Gaussian processes) to balance savings with performance objectives in the real world. If you have that framework, I could see this being extended to forms of memory other than local on-mobo RAM, like Flash drives, or actual RAM elsewhere in the datacenter: they just have worse performance characteristics than local RAM and get used less, is all.