[hs86]

While its deduplication feature clearly demands more memory, my understanding is that the ZFS ARC is treated by the kernel as a driver with a massive, persistent memory allocation that cannot be swapped out ("wired" pages). Unlike the regular file system cache, ARC's eviction is not directly managed by the kernel. Instead, ZFS itself is responsible for deciding when and how to shrink the ARC.

This can lead to problems under sudden memory pressure. Because the ARC does not immediately release memory when the system needs it, userland pages might get swapped out instead. This behavior is more noticeable on personal computers, where memory usage patterns are highly dynamic (applications are constantly being started, used, and closed). On servers, where workloads are more static and predictable, the impact is usually less severe.

I do wonder if this is also the case on Solaris or illumos, where there is no intermediate SPL between ZFS and the kernel. If so, I don't think that a hypothetical native integration of ZFS on macOS (or even Linux) would adopt the ARC in its current form.

[ryao]

The ZFS driver will release memory if the kernel requests it. The only integration level issue is that the free command does not show ARC as a buffer/cache, so it misrepresents reality, but as far as I know, this is an issue with caches used by various filesystems (e.g. extent caches). It is only obvious in the case of ZFS because the ARC can be so large. That is a feature, not a bug, since unused memory is wasted memory.

[pseudalopex]

> The ZFS driver will release memory if the kernel requests it.

Not fast enough always.

[dizhn]

Maz arc size is configurable and it does not need the mythical 1GB per TB to function well.

[netbsdusers]

Solaris achieved some kind of integration between the ARC and the VM subsystem as part of the VM2 project. I don't know any more details than that.

[ryao]

I assume that the VM2 project achieved something similar to the ABD changes that were done in OpenZFS. ABD replaced the use of SLAB buffers for ARC with lists of pages. The issue with SLAB buffers is that absurd amounts of work could be done to free memory, and a single long lived SLAB object would prevent any of it from mattering. Long lived slab objects caused excessive reclaim, slowed down the process of freeing enough memory to satisfy system needs and in some cases, prevented enough memory from being freed to satisfy system needs entirely. Switching to linked lists of pages fixed that since the memory being freed from ARC upon request would immediately become free rather than be deferred to when all of the objects in the SLAB had been freed.