Open krizhanovsky opened 5 months ago
The current MPMC lock-free queue is a nice engineering, but we always use per-cpu queues to push a work to a particular CPU, i.e. we use only MPSC case. MPMC queue has more synchronization overhead on atomics, so we should rework the queue to MPSC.
The task is easy to do since it's mostly about code simplification and improves performance for the core transport for network IO
The current MPMC lock-free queue is a nice engineering, but we always use per-cpu queues to push a work to a particular CPU, i.e. we use only MPSC case. MPMC queue has more synchronization overhead on atomics, so we should rework the queue to MPSC.