I know the approximation comes from an article used to estimate times in Python code, IIRC empirical. What I don't understand is why that formula ignores L2 accesses. I would expect them to produce a bigger hit than L1, as they are slower.
I'm asking because some code of mine produces a big (20%) increase in L2 accesses without changing RAM accesses or L1 accesses significantly as they're simply changing some small (two words) values in arguments to functions and returns from references to the pair to simply value copies.
I would expect that to inform a slowdown (I expect that to be slower than the original program), but instead I see speed estimation more or less unchanged, actually with a tiny negative number.
I know the approximation comes from an article used to estimate times in Python code, IIRC empirical. What I don't understand is why that formula ignores L2 accesses. I would expect them to produce a bigger hit than L1, as they are slower. I'm asking because some code of mine produces a big (20%) increase in L2 accesses without changing RAM accesses or L1 accesses significantly as they're simply changing some small (two words) values in arguments to functions and returns from references to the pair to simply value copies. I would expect that to inform a slowdown (I expect that to be slower than the original program), but instead I see speed estimation more or less unchanged, actually with a tiny negative number.
I haven't yet compared wall times tho.