On Tue, 17 Jul 2018 02:15:17 -0000, David Frank said: > inking. I'm checking out if the flag does what is is said to do-- I don't have > to call msync function, which would boost performance. Note that this can actually *kill* performance, because this means that the kernel has to flush to backing store every single time it notes a change, whereas if you use msync only at those points your software needs a sync point, it can do it at only those points.... Thought experiment: Imagine a workflow that needs to checkpoint every 1000 changes to the shared segment (for instance, if you've mapped an array with 1000 rows, do a for() loop across it incrementing one item, and checkpoint when they're all incremented). msync after the loop completes is one sync, while a worst-case using MAP_SYNC could result in a flush after every single increment (if the system is rescheduling the process over and over - for instance, if there's also another syscall inside the loop).
Attachment:
pgp3Pd4Jr8Xnv.pgp
Description: PGP signature
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@xxxxxxxxxxxxxxxxx https://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
