Messages in this thread | | | Subject | Re: [PATCH 7/8] sched: non-zero lag renice | From | Peter Zijlstra <> | Date | Fri, 24 Oct 2008 22:28:03 +0200 |
| |
On Fri, 2008-10-24 at 11:47 -0600, Chris Friesen wrote: > Peter Zijlstra wrote: > > > Then renicing, esp when lowering nice value (getting heavier), its possible > > to get into a starvation scenario. If you got too much runtime as a very > > light task, you get shot way far too the right, which means you'll have to > > wait for a long time in order to run again. > > > > If during that wait you get reniced down, fairness would suggest you get run > > earlier, because you deserve more time. > > > > This can be solved by scaling the vruntime so that we keep the real-time > > lag invariant. > > If we've already been shot way out to the right, presumably that would give us > a large real-time lag. If we renice to a lower nice level, wouldn't we want > to reduce the real-time lag rather than make it constant?
Ah, see but a 1ms real-time lag might be gigantic on weight=15, but nearly nothing on weight=88761.
1e6 * 1024/15 is massively larger than 1e6 * 1024/88761.
1000000*1024/15 = 68266666 1000000*1024/88761 = 11536
| |