On Wed, 2007-07-18 at 13:17 +0200, Giuseppe Ghibà wrote:
> Indeed not exists for PCI-e but Oden has spotted this PCI-X card
> (which is around 97$), based on marvell chipset:
> http://www.supermicro.com/products/accessories/addon/AoC-SAT2-MV8.cfm
Yes! There are several Broadcom and Marvell ASICs that are eight (8)
SATA (and now even some SAS) channels in a single chip, of which, there
are several vendors selling them for $100 boards. You do _not_ want an
on-board intelligence being an "intermediary" with software RAID, you
want _direct_ access to the SATA (or SAS) channels. Besides, they are
cheaper.
Justin Piszcz ha scritto:
> I am sure one of your questions is, well, why use SW RAID5 on the
> controller? Because SW RAID5 is usually much faster than HW RAID5,
> at least in my tests ...
Benchmarking direct disk access is a rather poor test of hardware RAID.
to truly evaluate if the off-load from the main system interconnect**
that a hardware RAID gives you, you need to benchmark your actual
_server_ application -- one with MD, the other with hardware RAID.
That's where you can tell if hardware RAID is going to buy you anything.
**NOTE: It's the system interconnect bottleneck that is of concern, not
so much the CPU. A modern, superscalar, multi-core x86-64 CPU can do
XORs in its sleep, with just MMX (won't even peg your CPU 10%). It's
the LOAD/STO push just to get the XOR that ties up your system
interconnect (often with your CPU only being 10% busy if your
application is not processor-bound, but I/O ;-) that's the problem.
It's been my experience that for web services, hardware RAID buys you
_little_, because you're more processor bound than I/O bound, so you
have cycles you can use while your interconnect is doing I/O processing.
But more on the database and file server side, I pair a hardware RAID
(e.g., AMCC/3Ware PPC400-based or Areca IOP/X-Scale-based) with a RX TOE
(_Receive_ TCP Off-load Engine) HBA "NIC" (e.g., LeWiz 4-port GbE PCIe),
which keeps the system inteconnect free for pushing the application data
(instead of doing what is, essentially, "programmed I/O" for the XOR
software RAID operation).
Again benchmark your server _application_, not the disk I/O.
--
Bryan J. Smith Professional, Technical Annoyance
mailto:b.j.smith@xxxxxxxx http://thebs413.blogspot.com
--------------------------------------------------------
Fission Power: An Inconvenient Solution
|