The problem with sbrk() is that if you allocate a few megabytes of memory + a tiny chunk, even after freeing everything apart from the tiny chunk near the "program break", nothing will be given back to the OS. If you use mmap() instead though, all pages can be given back apart from the one where the tiny chunk resides in. This makes for a tremendous difference sometimes.
You're talking about a resource reclamation issue. Unmapping a page is a clear signal to the kernel that the memory is unused and can be repurposed immediately. Otherwise, it needs to find and eject a page from memory using the VM system, which is more expensive (though I'd guess not a lot more expensive except in pathologically allocation-bound systems).
I'm not sure though that the VM system can reclaim memory in the heap allocated with sbrk(). At least I've never seen that before. Or do you mean "eject" as in swap out?
Obviously the valgrind advice is great. But I don't follow your point about mmap vs. sbrk regarding fragmentation. Fragmentation is a property of memory addresses, not the syscall used to allocate them. A big linear heap allocated with sbrk will fragment just as badly as the same heap allocated in a single block, even if it's allowed to have "holes" in it.
Maybe you're claiming that jemalloc deliberatly leaves gaps between allocations? That can help a little in practice, though the real effect is just to change the size of the fragmented blocks.