> This is not really accurate; the cfq scheduler reorders reads and writes > to have longer bursts, and these sometimes exceed the rate you set. I > understood this is mostly for separate rd/wr settings, or did you > reproduce it with "total" rates too?
I'll drop the comment. It's only with read/write settings and not with total. > > Also, it would be better to have a workaround for this. Perhaps we > could simply make the default value of max nonzero? In the old > throttling code the slice time is 0.1s, so perhaps we could see what > happens with max=0.1*avg. I already set max to 0.1*avg if needed in the code.