The sequential consistency model is used by default in atomic operations. However, it actually organizes work with atomic operations in such a way that all this work is performed as if in one thread, i.e. a global order is created and the processor can no longer perform simultaneous atomic operations. It turns out a consistent program. But we are moving to using low-level atomic operations and types to implement lock free, which is faster and more efficient than locks, but then why do we need sequential ordering, if it removes all parallelism at all? At the moment I present this as a tool for creating a prototype of the program, which later will be replaced in different places with other tools - ordering capture / release or weakened ordering. Those. in the final version, we have to get rid of consistent consistency, and so on. to get rid of unnecessary synchronization, which will achieve maximum performance on multiprocessor / multi-core systems?
|