If trivial allocations might cost you a microsecond then you can’t allocate at all on a latency-sensitive critical path and should probably get a better allocator. Also, be wary of power-of-two sized things because if they are aligned they are likely to have poor cache performance (too many things compete for the same set of cache lines).

It's not one allocation per element, as said previously it's a segmented vector. It works similarly to virtual memory: fixed-size segments are allocated for contiguous storage, while another parent array points to the segments. So appending/prepending either needs to append/prepend to the last/first segment or to allocate a new segment if full, which is what makes those operations more efficient than on a vector. Yes there is indirection, that's what segmented means. You also get stability of the memory location which is important for many data structures. The article even tells you how big the segments are in the two implementations being looked at.

Going back to HFT, a similar data structure that's definitely heavily used there is a pool, which has a configurable segment size (potentially variable) and doesn't have order, enabling faster deallocation with a freelist.