@ryanc@infosec.exchange the disadvantages of zram are well known (cpu overhead due to de/compression and the kernel struggling to free up unused cache space). A configured userspace oom killer clearly adds a lot in term of managing low memory conditions. If you have a machine you wish to run sets of virtual machines concurrently, I'd recommend doing big partitioned swap on nvme. But the reason I'm commenting: If you are doing the local AI thing and you feel like freeing up your nvme swap, you might want to look into repurposing the space into GPUDirectable KV cache: https://docs.lmcache.ai/mp/l2_storage/index.html Checked my neck; we can be friends :flan_thumbs: