lkml.org 
[lkml]   [2009]   [Jun]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 02/20] io-controller: Common flat fair queuing code in elevaotor layer
* Fabio Checconi <fchecconi@gmail.com> [2009-06-23 06:10:52]:

> > From: Vivek Goyal <vgoyal@redhat.com>
> > Date: Mon, Jun 22, 2009 10:43:37PM -0400
> >
> > On Mon, Jun 22, 2009 at 02:43:13PM +0200, Fabio Checconi wrote:
> >
> ...
> > > > Please help me understand this, we sort the tree by finish time, but
> > > > search by vtime, start_time. The worst case could easily be O(N),
> > > > right?
> > > >
> > >
> > > no, (again, the full answer is in the paper); the nice property of
> > > min_start is that it partitions the tree in two regions, one with
> > > eligible entities and one without any of them. once we know that
> > > there is one eligible entity (checking the min_start at the root)
> > > we can find the node i with min(F_i) subject to S_i < V walking down
> > > a single path from the root to the leftmost eligible entity. (we
> > > need to go to the right only if the subtree on the left contains
> > > no eligible entities at all.) since the RB tree is balanced this
> > > can be done in O(log N).
> > >
> >
> > Hi Fabio,
> >
> > When I go thorough the paper you mentioned above, they seem to have
> > sorted the tree based on eligible time (looks like equivalent of start
> > time) and then keep track of minimum deadline on each node (equivalnet of
> > finish time).
> >
> > We seem to be doing reverse in BFQ where we sort tree on finish time
> > and keep track of minimum start time on each node. Is there any specific
> > reason behind that?
> >
>
> Well... no specific reasons... I think that our implementation is easier
> to understand than the one of the paper, because it actually uses finish
> times as the ordering key, and min_start to quickly locate eligible
> subtrees, following the definition of the algorithm.
>

Is it still O(log N)?

> Moreover, if you look at the get_req() code in the paper, it needs a
> couple of loops to get to the result, while with our implementation
> we save the second loop.
>
> Our version is still correct, because it always moves to the left
> (towards smaller finish times), except when moving to the left would
> mean entering a non feasible subtree, in which case it moves to the
> right.
>
> Unfortunately I'm not aware of any paper describing a version of the
> algorithm more similar to the one we've implemented. Sorry for not
> having mentioned that difference in the comments nor anywhere else,
> it has been a long long time since I read the paper, and I must have
> forgotten about that.

/me needs to go read the paper in full.

--
Balbir


\
 
 \ /
  Last update: 2009-06-23 09:37    [from the cache]
©2003-2011 Jasper Spaans