[rayiner]

It really depends on the task. For an embedded device with a UI, where memory is at a premium, reference counting is probably the way to go.

But for pure throughput, especially on a multiprocessor machine, you really want to go with tracing garbage collector. In a multiprocessor system you have to keep caches coherent, and writes to the reference count have the effect of invalidating remote cache lines, increasing coherency traffic, etc.