Open dorlaor opened 11 years ago
I wonder if this one can be closed? Is the idea around tx coalescing still relevant?
To answer my question above I think this issue may still be relevant given Van Jacobson design only affects the RX path. On other hand, some comments and code in virtio-net driver suggest some coalescing logic may be implemented. Let us leave this issue in place.
Instead of kick the hypervisor for every tx invocation (can be more than one packet, depending on the tcp stack), ask the scheduler for a range timer between 0 - 200nsec. If there is an armed timer in that range, the thread doing the kick will get invoked. If there is no such armed timer, it will be invoked immediately since the timer arm is too expensive in virt.
This is low priority since at the moment there aren't much tx kicks and the stack structure will change as Van Jacobson design gets implemented.