I recognize this is quite possibly out of scope, insanely hard, or an insanely bad idea, etcetera.
What if there were options for selectively disabling specific processor functionality? For instance, there's occasional instances of things which have exceptionally swingy performance based on nothing more than "the specific address a function gets loaded at", which might have to do with branch prediction, cache misses, etcetera.
Being able to try things like "run this benchmark with L2 cache disabled" would be super helpful in debugging such things. I don't know whether it's practical or possible, these days, or would require additional help from a kernel or whatever.
I recognize this is quite possibly out of scope, insanely hard, or an insanely bad idea, etcetera.
What if there were options for selectively disabling specific processor functionality? For instance, there's occasional instances of things which have exceptionally swingy performance based on nothing more than "the specific address a function gets loaded at", which might have to do with branch prediction, cache misses, etcetera.
Being able to try things like "run this benchmark with L2 cache disabled" would be super helpful in debugging such things. I don't know whether it's practical or possible, these days, or would require additional help from a kernel or whatever.