We should consider: - throughput - latency - hot/cold cache - how many thread operation on the same location.