On Mon, Jun 04 2007, Rusty Russell wrote: > On Mon, 2007-06-04 at 15:43 +0200, Jens Axboe wrote: > > On Mon, Jun 04 2007, Carsten Otte wrote: > > > Jens Axboe wrote: > > > >On Fri, Jun 01 2007, Carsten Otte wrote: > > > >>With regard to compute power needed, almost none. The penalty is > > > >>latency, not overhead: A small request may sit on the request queue to > > > >>wait for other work to arrive until the queue gets unplugged. This > > > >>penality is compensated by the benefit of a good chance that more > > > >>requests will be merged during this time period. > > > >>If we have this method both in host and guest, we have twice the > > > >>penalty with no added benefit. > > > > > > > >I don't buy that argument. We can easily expose the unplug delay, so you > > > >can kill it at what ever level you want. Or you could just do it in the > > > >driver right now, but that is a bit hackish. > > > That would be preferable if the device driver can chose the unplug > > > delay, or even better it could be (guest)sysfs tuneable. > > > > Right. We probably should make it sysfs configurable in any case, right > > now it's a (somewhat) policy decision in the kernel with the delay and > > unplug depth. > > The danger is that it's just another knob noone knows how to use. > Perhaps simply setting it to 0 for the noop scheduler will cover all > known cases? Most people should not fiddle with it, the defaults are there for good reason. I can provide a blk_queue_unplug_thresholds(q, depth, delay) helper that you could use for the virtualized drivers, perhaps that would be better for that use? -- Jens Axboe _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/virtualization
