On Mon 19-12-22 03:16:33, 程垲涛 Chengkaitao Cheng wrote: > Hi Michal Hocko, > Looking forward to your reply. I am sorry, I do not have anything to add to my previous concerns. But let me summarize. I think your way of mixing per memcg protection with the per-process oom_score is very dubious. This is not an unfixable problem. All you need to do is the discount all processes in the same memcg equally. A bigger problem is, though, that I am not convinced the memory protection based interface is really viable. Based on experiences with the existing reclaim protection interface this is not really trivial interface to use. You either have to have a good overview of the working set size or you have to auto-tune it based on a feedback mechanism (e.g. PSI). Auto-tuning based on oom which should be a rare event is rather problematic I would say. All that being said I am not convinced that the interface is practically usable and you haven't really provided good examples to prove me wrong. -- Michal Hocko SUSE Labs
