Hello, Shaohua. On Wed, Apr 27, 2011 at 01:43:29PM +0800, Shaohua Li wrote: > > That would be a pathelogical case but, even then, after the change the > > number becomes much higher as it becomes a function of batch * > > num_updaters, right? > > I don't understand the difference between batch * num_updaters and batch > * num_cpus except preempt. So the only problem here is _add should have > preempt disabled? I agree preempt can make deviation worse. > except the preempt issue, are there other concerns against the atomic > convert? in the preempt disabled case, before/after the atomic convert > the deviation is the same (batch*num_cpus) Yes, with preemption disabled, I think the patheological worst case wouldn't be too different. > > I don't really worry about _sum performance. It's a quite slow path > > and most of the cost is from causing cacheline bounces anyway. That > > said, I don't see how the above would help the deviation problem. > > Let's say an updater reset per cpu counter but got preempted before > > updating the global counter. What differences does it make to check > > fbc->counter before & after like above? > > yes, this is a problem. Again I don't mind to disable preempt in _add. Okay, this communication failure isn't my fault. Please re-read what I wrote before, my concern wasn't primarily about pathological worst case - if that many concurrent updates are happening && the counter needs to be accurate, it can't even use atomic counter. It should be doing full exclusion around the counter and the associated operation _together_. I'm worried about sporadic erratic behavior happening regardless of update frequency and preemption would contribute but isn't necessary for that to happen. Thanks. -- tejun -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>