On Sat, 2016-01-30 at 09:59 +0100, James Bottomley wrote: > On Fri, 2016-01-29 at 17:40 -0500, Tejun Heo wrote: > > Hello, > > > > To implement meaningful IO resource accounting and distribution, > > it's > > necessary to know how much each IO or a group of IOs costs. With > > rotational devices, due to the inherent low performance and > > concurrency, direct IO cost estimation could be replaced with > > provisioning time slices - whatever compositions of IOs which can > > be > > performed in the same amount time are of the same cost. However, > > this > > approach can no longer be used for modern non-rotational > > high-performance devices with high level of concurrency. > > > > It'd be great to talk about the problem, how to work around the > > problem in the near future, and what solutions can be employed in > > the > > long term. > > Isn't the problem for non-rotational devices pretty similar to the > network traffic shaping one? In that model, there's a variety of > controllers (hfsc being the most frequently used one) which have > realtime (latency), linkshare (minimum bw) and upload (max bw) > parameters which we could just copy modulo a few changes. Probably out of scope for LSF/MM but this does beg the question of why traffic control in block is done by a cgroup but not in net ... I think the answer is "historical reasons", but if the net traffic classifiers at least were cgroups, we might get more sharing between net and block traffic control. James -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
