I will allow that GC can be fast, but you have to work hard and have appropriate language features to keep it "chunky". If you have millions of small traces to

I will allow that GC can be fast, but you have to work hard and have appropriate language features to keep it "chunky". If you have millions of small traces to follow, the n of your O(n) is going to be big. In contrast a manual collection algorithm can rely on user supplied metadata to free the whole dataset in one instant without having a tracing step. In apps that need to scale, designing data to be efficiently allocated and freed becomes crucial and the reference by default strategy really falls down.