Messages in this thread | | | Date | Mon, 30 Nov 2009 15:09:44 -0800 (PST) | From | David Rientjes <> | Subject | Re: [PATCH] oom_kill: use rss value instead of vm size for badness |
| |
On Fri, 27 Nov 2009, Andrea Arcangeli wrote:
> Ok I can see the fact by being dynamic and less predictable worries > you. The "second to last" tasks especially are going to be less > predictable, but the memory hog would normally end up accounting for > most of the memory and this should increase the badness delta between > the offending tasks (or tasks) and the innocent stuff, so making it > more reliable. The innocent stuff should be more and more paged out > from ram. So I tend to think it'll be much less likely to kill an > innocent task this way (as demonstrated in practice by your > measurement too), but it's true there's no guarantee it'll always do > the right thing, because it's a heuristic anyway, but even total_vm > doesn't provide guarantee unless your workload is stationary and your > badness scores are fixed and no virtual memory is ever allocated by > any task in the system and no new task are spawned. >
The purpose of /proc/pid/oom_adj is not always to polarize the heuristic for the task it represents, it allows userspace to define when a task is rogue. Working with total_vm as a baseline, it is simple to use the interface to tune the heuristic to prefer a certain task over another when its memory consumption goes beyond what is expected. With this interface, I can easily define when an application should be oom killed because it is using far more memory than expected. I can also disable oom killing completely for it, if necessary. Unless you have a consistent baseline for all tasks, the adjustment wouldn't contextually make any sense. Using rss does not allow users to statically define when a task is rogue and is dependent on the current state of memory at the time of oom.
I would support removing most of the other heuristics other than the baseline and the nodes intersection with mems_allowed to prefer tasks in the same cpuset, though, to make it easier to understand and tune.
| |