Paul Rubin wrote: > I haven't been keeping up with this stuff in recent years so I have a > worse concern. I don't know whether it's founded or not. Basically > in the past decade or so, memory has gotten 100x larger and cpu's have > gotten 100x faster, but memory is less than 10x faster once you're out > of the cpu cache. The mark phase of mark/sweep tends to have a very > random access pattern (at least for Lisp). In the old days that > wasn't so bad, since a random memory access took maybe a couple of cpu > cycles, but today it takes hundreds of cycles. So for applications > that use a lot of memory, simple mark/sweep may be a much worse dog > than it was in the Vax era, even if you don't mind the pauses.
You pay a price for CG one way or the other. In Python the price is spread out among a bunch of increment and decrement operations in the code. For mark and sweep the price is a big operation done less often. My understanding from reading about GC implementations is that reference counting can exhibit poor cache performance because decrementing a reference can lead to a chain of decrements on objects that can be laid out in memory in a fairly random fashion. Mark and sweep can be a better cache performer since it's memory access will tend to be more linear. -- http://mail.python.org/mailman/listinfo/python-list
