Liste des Groupes | Revenir à c arch |
On 9/23/2024 3:39 PM, John Dallman wrote:Affinity masks are nice, but iirc, they might be considered a sort of fairly "heavy" _hint_ to the underlying system, so to speak... In other words, it might not be 100% respected. Iirc I think I remember reading where a set affinity might not be held in 100% pure stone. Why does my mind go to Windows docs... ;^) It been a while. I remember as an older experiment I was pinning threads to cpus to try to get some sort of per cpu data in user space. I would day the thread local data on this pinned thread might as well be akin to per cpu data. What an experiment it was!In article <20240923105336.0000119b@yahoo.com>, already5chosen@yahoo.comThough...
(Michael S) wrote:
>mitchalsup@aol.com (MitchAlsup1) wrote:>When only 1 x86 would fit on a die, it really did not mater
much. I was at AMD when they were designing their memory
model.Why # of CPU cores on die is of particular importance?>
Because multi-core made multi-processor systems commonplace, and far more
software started using multiple threads.
>
If one added a scheduling constraint that two threads within a single process could not run at the same time on multiple cores, with each process assigned an affinity to a specific core...
Then memory consistency between cores would be less of an issue:
Threads would still behave as if there were TSO, on a weak-model CPU...
John
Les messages affichés proviennent d'usenet.