> By using swap, determinism is bye-bye as disk access, unlike RAM, is not deterministic.

No, because what's driving this change is the rise of NVMe rather than disk, which has consistent access times.

Consistent doesn't imply deterministic.

Most consumer DRAM is non-deterministic. A small fraction of users use tools to provide high/stronger deterministic guarantees but issues like bit flips and variable access latency still apply to DRAM although these issues are far smaller in timescale and scope.

None the less I'm assuming this set of optimizations is targeted not only at nvme storage but also 3D XPoint. Intel and Micron have done a lot of work in this space and promise to provide non-volatile, low latency, and high density in a single package. Micron has seemingly abandoned QuantX (their 3D XPoint) but Intel has recently made some huge strides in this area which, if adopted, will definitely push them to compete in this space.

Soon Intel Optane will be available in a DDR4-like package. It will still be closer to disk-like than RAM like and that is why I think this effort to push for large page swapping will be interesting. It being non-volatile also provides some cool possibilities for the future if this tech becomes more mainstream. Imagine being able to provision a portion of your swap that always contains the state of certain programs you care about or certain always-used libraries. Maybe even caching the entire OS in DRAM-like storage so that all boots read directly from your non-volatile swap.

Also, as a very important aside, this is from the patches mentioned in this article.

    From: Huang Ying <[email protected]>

Is RAM access time deterministic these days? What about cache levels and NUMA?

You're typically depending on disk access for your operating system to function and using swap-like mechanisms for most running executables (memory-mapped files can have pages loaded on demand and dropped), so your point is moot.