On 13/02/25 12:27, Christian Loehle wrote: > On 2/13/25 06:20, Juri Lelli wrote: > > On 12/02/25 19:22, Dietmar Eggemann wrote: > >> On 11/02/2025 11:42, Juri Lelli wrote: > > > > ... > > > >>> What about we actually ignore them consistently? We already do that for > >>> admission control, so maybe we can do that when rebuilding domains as > >>> well (until we find maybe a better way to deal with them). > >>> > >>> Does the following make any difference? > >> > >> It at least seems to solve the issue. And like you mentioned on irc, we > >> don't know the bw req of sugov anyway. > >> > >> So with this change we start with 'dl_bw->total_bw = 0' even w/ sugov tasks. > >> > >> dl_rq[0]: > >> .dl_nr_running : 0 > >> .dl_bw->bw : 996147 > >> .dl_bw->total_bw : 0 <-- ! > >> > >> IMHO, people who want to run serious DL can always check whether there > >> are already these infrastructural DL tasks or even avoid schedutil. > > > > It definitely not ideal and admittedly gross, but not worse than what we > > are doing already considering we ignore sugovs at AC and the current > > bandwidth allocation its there only to help with PI. So, duck tape. :/ > > > > A more proper way to work with this would entail coming up with sensible > > bandwidth allocation for sugovs, but that's most probably hardware > > specific, so I am not sure how we can make that general enough. > > > > Anyway, looks like Jon was still seeing the issue. I asked him to verify > > he is using all the proposed changes. Let's see what he reports. > > FWIW it also fixes my reproducer. > > I agree that dummy numbers for sugov bw is futile, but real bw numbers > also don't make a lot of sense (what if we exceed them? The system > won't be able to change frequency, i.e. might not be able to provide > bw for other DL tasks then either?). > I'm slightly worried about now allowing the last legal CPU for a sugov > cluster to offline, which would lead to a cluster still being active > but sugov DL unable to run anywhere. I can't reproduce this currently > though. Is this an issue in theory? Or am I missing something? Not sure I get what your worry is, sorry. In my understanding when the last cpu of a policy/cluster gets offlined the corresponding sugov kthread gets stopped as well (sugov_exit)?
