This is why generational GCs are a good idea. You can afford to waste RAM (proportionally) for early generations, because they're small. And for many use cases, this works out great; e.g. in server workloads, most of the allocations made during handling a request will die after the response has been sent. You can get great performance from having sufficient "waste" RAM for (say) 80th percentile total request allocation size times 80th percentile concurrent requests; but this is not normally a huge chunk of total heap size (which will usually be caches of various different kinds, which themselves benefit from a more manual deallocation strategy tied into cache expiration).