Re: [[PATCH, RESEND]] less AGs for single disks configs.

[David Chinner <dgc@sgi.com>]

On Mon, Nov 12, 2007 at 02:48:49PM +1100, xaiki@sgi.com wrote:
> From: Niv Sardi <xaiki@sgi.com>
> 
> get the underlying structure with get_subvol_stripe_wrapper(),
> and pass sunit | swidth as an argument to calc_default_ag_geometry().
> 
> if it is set, get the AG sizes bigger.
> 
> this also cleans up a typo:
> -       } else if (daflag)      /* User-specified AG size */
> +       } else if (daflag)      /* User-specified AG count */

No need to mention that you are cleaning up a typo in the description ;)

> ---
>  xfsprogs/mkfs/xfs_mkfs.c |   18 ++++++++++++------
>  1 files changed, 12 insertions(+), 6 deletions(-)
> 
> diff --git a/xfsprogs/mkfs/xfs_mkfs.c b/xfsprogs/mkfs/xfs_mkfs.c
> index 78c2c77..4cf9975 100644
> --- a/xfsprogs/mkfs/xfs_mkfs.c
> +++ b/xfsprogs/mkfs/xfs_mkfs.c
> @@ -393,6 +393,7 @@ void
>  calc_default_ag_geometry(
>  	int		blocklog,
>  	__uint64_t	dblocks,
> +	int		multidisk,
>  	__uint64_t	*agsize,
>  	__uint64_t	*agcount)
>  {
> @@ -428,12 +429,13 @@ calc_default_ag_geometry(
>  	 *
>  	 * This scales us up smoothly between min/max AG sizes.
>  	 */
> +
>  	if (dblocks > GIGABYTES(512, blocklog))
> -		shift = 5;
> +		shift = 5 - (multidisk == 0);
>  	else if (dblocks > GIGABYTES(8, blocklog))
> -		shift = 4;
> +		shift = 4 - (multidisk == 0);
>  	else if (dblocks >= MEGABYTES(128, blocklog))
> -		shift = 3;
> +		shift = 3 - (multidisk == 0);
>  	else
>  		ASSERT(0);

Ok, so now we end up with half the number of allocation groups
at these different sizes. That's not exactly what I had in mind.
basically, what you've done works out as:

	> 512GB  old = 32 AGs, new = 16AGs
	> 8 GB   old = 16 AGs, new = 8AGs
	> 128MB  old = 8 AGs, new = 4AGs

on an 8Gb filesystem we still get 8 AGs, which is far too many.
on a 750GB disk, we still get 16AGs, which to far too many.

A single spindle, regardless of it's size, will have similar
seek characteristics so scaling the number of AGs with size
is the wrong thing to do - you don't get better parallelism
out of a single spindle, just more seeks and lower performance.
hence keeping the number of AGs fixed up to the point where
the AG size tops out (i.e. 4TB) seems like a better scaling
factor to me. i.e. something like:


	if (!multidisk) {
		if (dblocks >= TERABYTES(4, blocklog)) {
			blocks = XFS_AG_MAX_BLOCKS(blocklog);
			goto done;
		}
		agcount = 4;
		/* work out ag size here */
		goto done;
	}

I'd also like to see some test results showing the mkfs output
for the different configurations to confirm it works correctly
(i.e. that the corner cases work correctly).

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group