Hello, Tao Ma. On Sat, Sep 01, 2012 at 09:58:43PM +0800, Tao Ma wrote: > Vivek and I have talked about its usage in my first try. See the thread > here. https://lkml.org/lkml/2012/5/22/81 > And I am OK to say it again here. In our case, we use flashcache as a > block device and the bad thing is that flashcache is a bio-based dm > target and we can't use block io controller here to control the weight > of different cgroups. So io throttle is chosen. But as io throttle can > only set a hard upper limit for different instances, it makes the > control not flexible enough. Say with io controller, if there is no > requests form the cgroup with weight 1000, a cgroup with 500 can use the > whole bandwidth of the underlying device. But if we set 1000 iops for > cgroup A and 500 iops for cgroup B in io throttle, cgroup B can't exceed > its limit even if cgroup A has no request pending. So if we can export > the io_queued information out to the system admin, they can write some > daemon and in the above case, increase the upper limit of cgroup B to > some number say 1000. It helps us to utilize the device more > efficiently. Does it make sense to you?
Somewhat, in a pretty twisted way. :P > > Adding throttle.io_queued could be a bit more consistent? > > sorry, I don't know what is your meaning here. You mean some codes like > blkg_rwstat_add(&stats_cpu->throttle.io_queude, rw, 1)? So, there already is io_dispatched, so if you have io_queued, you can read the two and calculate the difference from userland (reading io_queued first would probably be better to avoid triggering the throttled condition spuriously). That way, you don't have to worry about synchronizing stats across cpus and it's a simple addition of a stat conter. Thanks. -- tejun -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
