Big vectors inherently avoid sub-page fragmentation because they're allocated directly via mmap. The only thing that matters is having unreserved address space, and 64bit gives you a lot.

What you say is true for a worst case load of millions of vectors that are larger than half of the mmap threshold size (typically roughly page size or a small multiple of it). So this might just be a semantic disagreement, but I don't think of millions of 3KB vectors as 'big' vs a smaller number of MB or GB vectors.