[Pc_Support] Intel ICH7R and nVidia MCP-04 at RAID-5 = 15MBps (yes, it's FRAID)

[Bryan J. Smith]

I'm going to post this to LEAP as well, because just today I had to
explain FRAID (fake RAID) yet again.

- What is FRAID (Fake RAID)?  And why is it popular?

FRAID is _software_ RAID done in the vendor's driver for the OS.  The
reason it is popular is because it requires *0* additional hardware.
The vendor takes a _dumb_ ATA bus arbitrator, and adds a little extra
logic into the 16-bit BIOS Int13h Disk Services.  You think you get
hardware RAID, but you don't.  All RAID functionality is 100% driven by
your CPU.

You see, once the OS boots, the little extra 16-bit routines are
_useless_ and a driver is required.  But instead of letting the OS
handle the disk organization, it has already been fixed in the 16-bit
Int13h Disk Services.  So it's up to the software driver, with some 3rd
party licensed code, to do the RAID.  It's that 3rd party licensed code
that is proprietary and will never have a GPL release.

A _real_, intelligent hardware RAID card has on-board intelligence with
firmware that drives all RAID functionality.  The OS driver only talks
to that intelligence, not the disks themselves, for block transfers.
The on-board intelligence does rebuilds, XORs, etc...

This is important to note because RAID-1 and, especially, RAID-5 is
extremely poor in software.  Not because of the XOR operation, but
because of the redundant streams of data that much pass through the CPU
during a write.  Instead of a simple block transfer from memory directly
to I/O, and the intelligence doing any mirroring/parity, it gets jammed
through the CPU first, which adds _massive_ overhead.

- So what?  Why should I care?

Well, the comment came up today that RAID-5 is always hardware.  Not so!
The new Intel I/O Controller Hub version 7 RAID (ICH7R) and nVidia Media
and Communication Processor (MCP-04) chips now offer so-called RAID-5
support.  They are nothing more than FRAID, with the CPU still doing
_all_ work.

If you still don't believe how _bad_ software RAID-5, you need to check
out these charts from GamePC ...

The first one shows _write_ performance of RAID-5:  
  http://www.gamepc.com/labs/view_content.asp?id=raid505&page=9  

You'll note the ICH7R and MCP-04 at the _bottom_.  Even the Areca
ARC-1110 (superscalar XScale Intel I/O Processor 331, IOP331) in a 32-
bit PCI slot is much faster, as well as the Broadcom (which is partially
FRAID, long story).  In reality, these tests were _not_ good because
they only use 1GB writes maximum -- a bigger write would have been
better (let alone with _multiple_ client loads like a server would
have).

Now this next one shows where hardware RAID rules (second graph):  
  http://www.gamepc.com/labs/view_content.asp?id=raid505&page=10  

As the disk cache fills up, only the Areca performs (in a 64-bit slot --
much be an issue with it in a 32-bit slot).  Even the Broadcom quickly
gets overloaded with pushing too much data up the CPU interconnect after
a few MBs.  Again, I'd really like to see multi-GB files used, which
would be much better at showing off the real performance.

And in any case, 3Ware _rules_ the I/O queuing -- which is what servers
are all about.  No one has been able to show an ATA/SATA solution that
bests it at raw queuing -- and it really challenges even most SCSI RAID
solutions as well (all but the top-end IOP331-333 designs).


-- 
Bryan J. Smith     b.j.smith at ieee.org     http://thebs413.blogspot.com
----------------------------------------------------------------------
The best things in life are NOT free - which is why life is easiest if
you save all the bills until you can share them with the perfect woman