How about take the basic token bucket algorithm first, and consider optimize options later? Such as: - limit the idle reserve - change the static depth of token bucket - calculate the depth of token bucket according to configured bps rate - export the depth of token bucket to user space On Wed, Oct 16, 2013 at 11:53 PM, Tejun Heo <tj@xxxxxxxxxx> wrote: > Hello, > > On Wed, Oct 16, 2013 at 10:14:06AM -0400, Vivek Goyal wrote: >> - First of all, if you think that a group is entitiled for tokens even >> when it is not doing IO, then why are you truncating the tokens after >> dispatch of a BIO. >> >> - Second in general it does not seem right that a group is entitiled to >> tokens even when no IO is happening or group is not backlogged. That >> would mean a group will not do IO for 10 hours and then be entitiled >> to those tokens suddenly after 10 hours with a huge burst. >> >> So I think you also agree that a group should not be entitiled to >> tokens when group is not backlogged and that's why you seem to be >> truncating extra tokens after dispatch of a BIO. If that's the case, >> then even for first BIO, ideally a group should not be given tokens >> for idle time. > > Without going into details, having token reserve is an important part > of token based implementation. The large the reserve could be > debatable but that's what provides "smoothing" of allocation. e.g. if > you trim bucket as soon as the queue becomes empty, a queue with > sequential access pattern can easily get disadvantaged. Another way > to look at it is to consider as though the IO has been issued some > time before than actual and waited for the token - it is the same to > external observers. > > So, while how large the reserve should be is definitely debatable, > bucket scheduling *needs* idle reserve. > > Thanks. > > -- > tejun -- best regards Hong Zhiguo -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
