Swap is expensive compared to releasing unused memory back to the OS. The reason is that you spend memory and disk bandwidth writing “unused” data to disk. And that data could very well be unused RAM just sitting around in a memory allocator, which is effectively useless memory that you’re swapping because the allocator didn’t release it.

Zswap is always performance increasing. Swap to disk can be performance degrading (good implementations generally are not unless your working set is larger than your memory and you’re in thrashing) and certainly expensive $$ wise in that it wears out your SSD faster.

You seem to be thinking I’m arguing in absolute terms where all I’m saying is that swapping is a more expensive technique to try to reclaim that unused RAM vs the memory allocator doing it. It can be a useful low-effort technique, but it’s very coarse and more of a stop gap to recover inefficiencies in the system at a global level. Global inefficiency mitigation is generally not as optimally effective as more localized approaches.

Consider also that before the OS starts swapping, the OS is going to start purging disk caches since those are free for it to reload (executable code backed by file, page caches etc). These are second order system effects that are hard to reason about abstractly when you have a greedy view of “my application is the only one that matters”. This means that before you even hit swap, your large dark matter of dirty memory sitting in your allocator is making your disk accessed slower. And the kernel’s swap doesn’t distinguish working set memory from allocator so you’re hoping inherent temporal and spatial locality patterns interplay well so that you’re not trying to hand out an allocation for a swapped out block too frequently.