Re: raw vs block speed difference?

["Aaron J. Grier" <agrier%poofygoof.com@localhost>]

On Fri, Aug 27, 2010 at 08:29:34AM -0600, Michael L. Hitch wrote:
> On Thu, 26 Aug 2010, Aaron J. Grier wrote:
> 
> ># newfs -b 64k sd3d
> >/dev/rsd3d: 17501.8MB (35843664 sectors) block size 65536, fragment size 
> >8192
> >       using 6 cylinder groups of 2917.00MB, 46672 blks, 93184 inodes.
> 
> I'm not sure if newfs updates the file system block sizes in the disk
> label anymore, so the label information may not reflect the actual
> sizes.  In addition, I think the size used is actually the fragment
> size, which is 8K in this case.

On Fri, Aug 27, 2010 at 04:23:10AM -0400, jmitchel%bigjar.com@localhost wrote:
> If you mount the filesystem and do a dd to a file on the disk do you
> see the same throughput? I'm wondering if newfs has some overhead in
> it. In other words does:
> 
> mount /dev/rsd3d /mnt
> dd if=/dev/zero of=/mnt/temp count=100000 bs=16384
> 
> yield a different throughput result?

this was a bit eye-opening.

# newfs -b 16k sd0b
[...mount, cd...]
# dd if=/dev/zero of=bigfile count=130k bs=16k 
133120+0 records in
133120+0 records out
2181038080 bytes transferred in 42.100 secs (51806130 bytes/sec)
# dd if=bigfile of=/dev/null bs=16k           
133120+0 records in
133120+0 records out
2181038080 bytes transferred in 41.957 secs (51982698 bytes/sec)

this is close to the raw device speed.  (64k blocks gives similar
numbers.)

On Fri, Aug 27, 2010 at 08:29:34AM -0600, Michael L. Hitch wrote:
> You can run iostat during the transfer and see what the actual read
> size of the disk is.

      tty              sd0             CPU
 tin tout  KB/t  t/s  MB/s  us ni sy in id
[ writes ]
   0   18 63.84  792 49.38   0  0  6  0 94
   0   18 63.95  805 50.25   0  0  6  0 94
   0   18 63.75  783 48.73   0  0  5  0 94
[ reads ]
   0   18 63.85  822 51.24   0  0  5  0 95
   0   18 63.89  821 51.19   0  0  5  0 95
   0   18 63.87  820 51.17   0  0  5  0 95

looks like things are being amortized?

the moral of the story is that a dd of a block device is no indicator of
filesystem performance, so I apparently have nothing to be concerned
about.

back to RAIDframe tuning.  ;)

-- 
  Aaron J. Grier | "Not your ordinary poofy goof." | 
agrier%poofygoof.com@localhost