[linux-lvm] poor read performance on rbd+LVM, LVM overload

[Sage Weil]

On Mon, 21 Oct 2013, Mike Snitzer wrote:
> On Mon, Oct 21 2013 at 10:11am -0400,
> Christoph Hellwig <hch at infradead.org> wrote:
> 
> > On Sun, Oct 20, 2013 at 08:58:58PM -0700, Sage Weil wrote:
> > > It looks like without LVM we're getting 128KB requests (which IIRC is 
> > > typical), but with LVM it's only 4KB.  Unfortunately my memory is a bit 
> > > fuzzy here, but I seem to recall a property on the request_queue or device 
> > > that affected this.  RBD is currently doing
> > 
> > Unfortunately most device mapper modules still split all I/O into 4k
> > chunks before handling them.  They rely on the elevator to merge them
> > back together down the line, which isn't overly efficient but should at
> > least provide larger segments for the common cases.
> 
> It isn't DM that splits the IO into 4K chunks; it is the VM subsystem
> no?  Unless care is taken to assemble larger bios (higher up the IO
> stack, e.g. in XFS), all buffered IO will come to bio-based DM targets
> in $PAGE_SIZE granularity.
> 
> I would expect direct IO to before better here because it will make use
> of bio_add_page to build up larger IOs.

I do know that we regularly see 128 KB requests when we put XFS (or 
whatever else) directly on top of /dev/rbd*.

> Taking a step back, the rbd driver is exposing both the minimum_io_size
> and optimal_io_size as 4M.  This symmetry will cause XFS to _not_ detect
> the exposed limits as striping.  Therefore, AFAIK, XFS won't take steps
> to respect the limits when it assembles its bios (via bio_add_page).
> 
> Sage, any reason why you don't use traditional raid geomtry based IO
> limits?, e.g.:
> 
> minimum_io_size = raid chunk size
> optimal_io_size = raid chunk size * N stripes (aka full stripe)

We are... by default we stripe 4M chunks across 4M objects.  You're 
suggesting it would actually help to advertise a smaller minimim_io_size 
(say, 1MB)?  This could easily be made tunable.

sage