Re: [9fans] Re: NIX this morning

[David Leimbach via 9fans]

Sorry for the stream of consciousness that follows (kinda or I wouldn’t post it) Generally, on time slicing systems you cannot simply assign threads to cores 1:1 and expect to get the best efficiency for a long running process.  Usually your wall clock times end up being reduced by oversubscribing CPU cores. With AC’s hopefully one can get more predictable results but I do wonder about how system calls generally work as a traditional system takes a voluntary context switch to process them. If one had the ability to overlap communication (syscalls) with computation to a high degree with a given program, something like io_uring perhaps for async calls to the kernel, one could potentially get rid if some of the “voluntary” context switching as well. I can see how the new semaphore implementation may help here. Interesting … Sent from my iPhone On Feb 9, 2025, at 11:18 AM, hiro <23h...@gmail.com> wrote:  very interesting. is this interference metric normalized taking the efficiency of the system into account? e.g. if on one platform 2x as much work is done (for example because the core frequency is doubled), i wouldn't be surprised if absolute interference doubles, but in relation to the work being achieved i would assume the interference metric could be much less then double. another example: if there's a very low scheduling granularity, and very short jobs, interference noise floor might be higher, bec. the workload itself has much less impact in comparison to the idle periods. 9fans / 9fans / see discussions + participants + delivery options Permalink