* Jens Axboe <axboe@xxxxxxxxx> wrote: > - With the former note, the app either needs to opt in (and hence > willingly sacrifice CPU cycles of its scheduling slice) or it needs to > be nicer in when it gives up and goes back to irq driven IO. The scheduler could look at sleep latency averages of the task in question - we measure that already in most cases. If the 'average sleep latency' is below a certain threshold, the scheduler, if it sees that the CPU is about to go idle, could delay doing the context switch and do "light idle-polling", for say twice the length of the expected sleep latency - assuming the CPU is otherwise idle - before it really schedules away the task and the CPU goes idle. This would still require an IRQ and a wakeup to be taken, but would avoid the context switch. Yet I have an ungood feeling about depending on actual latency values so explicitly. There will have to be a cutoff value, and if a workload is just below or just above that threshold then behavior will change markedly. Such schemes rarely worked out nicely in the past. [Might still be worth trying it.] Couldn't the block device driver itself estimate the expected latency of IO completion and simply poll if that's expected to be very short [such as there's only a single outstanding IO to a RAM backed device]? IO drivers doing some polling and waiting in the microseconds range isnt overly controversial. I'd even do that if the CPU is busy otherwise: the task should see a proportional slowdown as load increases, with no change in IO queueing behavior. Thanks, Ingo -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
