In this Tuning blog post, Intel’s Shannon Cepeda recently was told by a customer that he had introduced a successful optimization to a hot function within his application, but didn’t see as much improvement in the overall application as he’d hoped. Cepeda writes that this isn’t uncommon in the iterative process of performance tuning, and provides an explanation of why it usually it happens.
To read the full blog post, click here.
Performance improvement estimates using Amdahl's law only apply if the user does not increase computing demands given more computing power. I got this link from the newsletter "Go Parallel: Translating Multicore Power into Application Performance". So let me assume that you are vectorizing on more than one core. So I'm using my 3930K hex core running at 3.8 GHz. In calculating speed-up I must consider memory latency in relation to my processor performance. The latency in terms of processor cycles is constantly increasing. In my speed-up calculations I must account for the extent to which improved processor performance will translate into improved computational performance.