Re: testing raid 0

["Paul M. Hirsch" <elektrosatan_(_at_)_voltagenoir_(_dot_)_org>]

On Wed, Aug 01, 2001 at 08:55:17AM -0700, Chris Palmer wrote:
> 
> Why is RAIDframe slower than a single disk?

It adds overhead.  If tuned properly, it will be faster than a single disk
by a long shot at reading.  Depending on hardware, writing may be faster
as well (though my results below don't show it :)

Here are some Bonnie runs on a pair of 10K RPM 9GB SCSI drives. The first
part of the "Name" for each test run represents the partition type: single
disk (sin), a RAIDFrame RAID0 set (rai), and a CCD set (guess?).  The first
number in the name is the filesystem block size, and the second is the stripe
size.  All tests were done using identical disks with equal base partitions.

              -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Name	   MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU  /sec %CPU
sin8      512 23270 61.6 22863 18.4  4854  3.0 27078 82.1 27120 10.4 145.7  0.7
sin32     512 23770 60.6 23681 19.6  6365  3.5 27071 80.2 27144  9.4 132.9  0.7
rai8-256  512 15816 38.8 13598  9.7  7358  8.8 29127 90.7 43145 36.1 166.5  1.9 
rai32-256 512 21185 50.1 18456 14.3 11623 12.6 28247 86.1 44035 33.5 145.8  1.6
ccd8-256  512 36239 87.2 37418 30.9 11207 11.5 30763 93.2 46197 35.2 159.8  1.3
ccd32-256 512 30523 75.0 30814 27.5  7844  7.6 27888 83.2 41158 29.2 146.2  1.2

As you can see, RAIDFrame beats the single spindle in block read performance,
but tanks at writes at both 8k and 32k block sizes.  Its seek performance
with 8k blocks trounces the single disk.  CCD beats both single and RAIDFrame,
especially on the 8k per block test.  I would recommend CCD striping for
general use, and maybe RAID0 if seek performance is most important.
As always, test with the application you will be using, or at least with
something stresses the same things.  Test, tweak, test. Don't just use the
stripe settings I show, either.  256 was the right size on this machine, 
but was all wrong on another pair of boxes I have.

-Paul

-- 
/Paul M. Hirsch              /
/elektrosatan_(_at_)_voltagenoir_(_dot_)_org/
/GPGPGPkeyID: 0xD11A250E     /