Re: [TCLUG:5717] raid-5 on one hard drive

[Dan Mack <mack@sgi.com>]

On Fri, 30 Apr 1999, Ben Luey wrote:

> Apr 29 20:20:40 pclueyb kernel: raid5: using high-speed MMX checksum
> routine               
> 
> And listed different speeds for MMX P-II, MMX-P and 32bit and 8big (P-II
> being the fastest). I though raid5 combined hard drives for redundany and
> speed? How can raid5 work on my computer with one hard drive? Should I
> recompile my kernel (not rh stock) to have raid 5 support? I have a k6-2
> is this matters (is it really using MMX?)

I would not bother re-compiling your kernel, unless you are low on memory
or just want to do it for fun. The md driver is just letting you know
which routine it *would* use if your were to configure and setup any raid5
volumes. Raid5 performance is fairly sensitive to how fast one can compute
the XOR of a set of values.  On my system:

  | md driver 0.90.0 MAX_MD_DEVS=256, MAX_REAL=12
  | raid5: measuring checksumming speed
  | raid5: using high-speed MMX checksum routine
  |    pII_mmx   :   638.175 MB/sec
  |    p5_mmx    :   623.697 MB/sec
  |    8regs     :   460.629 MB/sec
  |    32regs    :   330.327 MB/sec
  | using fastest function: pII_mmx (638.175 MB/sec)

md will be using the pII_mmx routine because it is the fastest that my
machine can do the checksumming...

If you aren't going to do any raid5, then you can ignore this message.

Also, see /usr/src/linux/drivers/block/raid5.c and
/usr/src/linux/drivers/xor.c for more information, namely:

<from /usr/src/linux/drivers/block/raid5.c>

/*
 * we use the 'XOR function template' to register multiple xor
 * functions runtime. The kernel measures their speed upon bootup
 * and decides which one to use. (compile-time registration is
 * not enough as certain CPU features like MMX can only be detected
 * runtime)
 *
 * this architecture makes it pretty easy to add new routines
 * that are faster on certain CPUs, without killing other CPU's
 * 'native' routine. Although the current routines are belived
 * to be the physically fastest ones on all CPUs tested, but
 * feel free to prove me wrong and add yet another routine =B-)
 * --mingo
 */

- -- -- --- --- --- ---- ---- ----- ----- ----- ----- ------
Dan Mack (mack@sgi.com)      <http://thistle.cray.com/~mack>
SGI -----  --     --   -----   -- -   ----  --  - -- --- ---