For 2) last time I looked I couldn't find a way to have a memory contract with the kernel that allowed for a given allocation to be evicted at will by the kernel while also telling the application that has happened. This would be ideal for apps that can grow their cache unbounded (e.g., photo processing apps with each photo, browsers with each tab) but can also regenerate that cache at will. You end up having tunables for how much cache to hold in RAM which is never correct as depending on what else the user is doing there may be a lot or very little free RAM.