On Fri, May 29, 2009 at 12:06:03PM -0700, Nauman Rafique wrote: > On Fri, May 29, 2009 at 9:57 AM, Fabio Checconi <fchecconi@xxxxxxxxx> wrote: > >> From: Vivek Goyal <vgoyal@xxxxxxxxxx> > >> Date: Fri, May 29, 2009 12:06:10PM -0400 > >> > >> On Thu, May 28, 2009 at 12:41:27PM -0700, Nauman Rafique wrote: > > ... > >> > I have some concerns about the new preemption logic. > >> > >> Actually we need a more proper definition of in-class preemption. Across > >> class preemption means that RT class always gets to run first. > >> > >> What does in-class preemption mean? If I look at the current CFQ code, > >> it does look like that preempting process will gain share. It is always > >> added to the front of the tree with "rb_key=0" and that means, this new > >> queue will get fresh time slice (even if it got time slice very recently). > >> > >> Currently I have just tried to make the behavior same as CFQ to reduce > >> the possiblility of regressions. That's a different thing that we can > >> discuss what should be the exact behavior in case of in-class preemption > >> and first it needs to be fixed in CFQ, if current behavior is an issue. > >> > >> On the other hand, I am not sure if previous bfq preemption logic was > >> working. We were checking if the new request belonged to the queue which > >> will be served next, then preempt the existing queue. While looking > >> for the next queue, I think we did not consider the current active > >> entity (as it was not on the tree). So after expiry of the current > >> queue, it might get selected next if it has not got its share. So there > >> was no point in preempting the queue. If queue already got its share, then > >> anyway the next queue will be selected next and there is no point in > >> preempting the current queue. > >> > > > > BFQ had no preemption logic, as far as I know; it simply was not > > preemptive, and the guarantees it provided took that into account. > > > > I don't know what is the best way to introduce a CFQ-like preemption logic > > into the wf2q+ code; for sure anything that does not schedule according > > to the algorithm's timestamps is a good candidate to break the guarantees > > the scheduler can provide, making it an extremely complex way to get > > the same worst-case delays of a (much simpler) round-robin scheduler. > > > > What you guys think of my suggestion of handling preemption? > Basically, we don't modify the start/finish tags, so overall the > fairness properties should not be broken. But in short term, we still > allow preemption and let one queue jump another. It sounded complicated from the description of it. I would prefer either we get rid of in-class preemtion thing completely or do in-class preemtption at the cost of gaining share, like cfq does. In fact, to begin with, I prefer to be as close as possible to CFQ and then change things selectively one piece at a time so that we can analyze the impact well. Thanks Vivek -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
