[Devel] Re: [PATCH 9/9] ext3: do not throttle metadata and journal IO

[Balbir Singh]

* Theodore Tso <tytso at mit.edu> [2009-04-21 13:46:20]:

> On Tue, Apr 21, 2009 at 10:53:17PM +0530, Balbir Singh wrote:
> > Coming to the dirty page tracking issue, the issue that is being
> > brought about is the same issue that we have shared page accounting. I
> > am working on estimates for shared page accounting and it should be
> > possible to extend it to dirty shared page accounting. Using the
> > shared ratios for decisions might be a better strategy.
> 
> It's the same issue, but again, consider the use case where the
> readers and the writers are in different cgroups.  This can happen
> quite often in database workloads, where you might have many readers,
> and a single process doing the database update.  Or the case where you
> have one process in one cgroup doing a tail -f of some log file, and
> another process doing writing to the log file.
> 

That would be true in general, but only the process writing to the
file will dirty it. So dirty already accounts for the read/write
split. I'd assume that the cost is only for the dirty page, since we
do IO only on write in this case, unless I am missing something very
obvious.

> Using a shared ratio is certainly better than charging 100% of the
> write to whichever unfortunate process happened to first read the
> page, but it will still not be terribly accurate.  A lot really
> depends on how you expect these cgroup limits will be used, and what
> the requirements actually will be with respect to accuracy.  If the
> requirements for accuracy are different for RSS tracking and dirty
> page tracking --- which could easily be the case, since memory is
> usually much cheaper than I/O bandwidth, and there is generally far
> more clean memory pages than there are dirty memory pages, so a small
> numberical error in dirty page accounting translates to a much larger
> percentage error than read-only RSS page accounting --- it may make
> sense to use different mechanisms for tracking the two, given the
> different requirements and differring overhead implications.
>
> Anyway, something for you to think about.

Yep, but I would recommend using the controller we have, if the
overheads span out to be too large for IO, we think about
alternatives.

-- 
	Balbir
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers