[Novalug] best hard disk setup for home file server?
Bryan J. Smith
b.j.smith at ieee.org
Wed Oct 14 16:41:26 EDT 2009
Of course. A full card with a real microcontroller core + XOR ASIC
peripherials to do XORs in-line (without the full LOAD-XOR-STO SIMD
approach) are not cheap, let alone have their capacitors, SRAM/DRAM
and optional battery-backup units (BBUs, if DRAM is used).
There are several options, but most boil down to ...
- AMCC/3Ware PowerPC 400 + ASIC Storage Switches (9550SX-9690SA series)
- Marvell/Areca XScale (superscalar ARM) IOP31x-34x Products
_Never_ purchase an i960 microcontroller IOP30x product. They are '90s
technology and major slouches. But people still seem to buy them, and it took
years before Promise's SuperTrak finally adopted XScale IOP31x+ logic
(well after Areca started stealing marketshare). They can't even do 50MBps
Anything 3Ware 7000-9500S is going to be 64-bit ASIC+SRAM only (with
the 9500S adding DRAM, but still ASIC-based). They are great, non-blocking
designs for RAID-0/1/10, but there's not enough SRAM for RAID-5 buffering
and no microcontroller in the 9500S. They are 64-bit PCI/PCI-X only.
[ NOTE: With that said, the older 8506 and 9500S still make great, non-blocking
RAID-10 cards with 4-12 disks, or even the 8006-2LP for just simple RAID-1
mirroring. But they really require their own 64-bit PCI/PCI-X channel to be
effective and not block other I/O on a shared PCI bus. Just don't throw RAID-5
at them or expect the ultimate of performance these days. ]
LSI Logic just bought AMCC and is likely going to pollute the "pure hardware"
design of 3Ware. I.e., LSI has been toting that future designs will be sold with
their software management stack value-add. That really puts into question
3Ware's long-standing GPL-only support, releasing volume layout and specs
so DeviceMapper can read their volumes and SMARTtools can interface with
individual drives, etc...
Any anything Areca is based on the former Intel XScale which Intel sold to
Marvell, so it has a questionable future. There was much talk of the Atom cores
adding ASICs for hardware RAID like the XScale IOP31x-34x has, but so far
Intel hasn't moved it's focus away from Netbooks on the Atom. I haven't seen
any products at all based on an Atom core with the required ASIC.
I.e., Intel has been promising for almost a decade that they'd integrate their
IOP logic/ASICs directly in their ICH/ESB (I/O Controller Hub / Enhanced
South Bridge) so you got hardware assisted RAID "for free" in mainboards.
And they haven't moved jack on that, and we keep getting Fake RAID
(absolute 0 RAID assist, only a 16-bit "trick BIOS" for setup and a 100%
software RAID "driver" for the 32/64-bit OSes -- DeviceMapper is now used
in Linux to remap volume layout for FRAID).
XOR ASICs are the reasons why people choose hardware RAID. General
microcontrollers and microprocessors are wholly inefficient at the LOAD-XOR-
STO flow (even with SIMD) and XOR ASICs on the datapath itself allow XORs
to be calculated in-line with the bus. Without them I'd rather just use software
RAID. I.e., FRAID-5 typically cannot break 15MBps writes.
In fact, what I common advocate "on-the-cheap" is a FRAID-1 (two mirrored
disks) just for system volumes (/, /tmp, /var, swap, etc...) and then software
RAID-6 for everything else (anything largely read, like /usr, plus data). The
former allows boot failover (with DeviceMapper) and the latter comes up
Software RAID is quite feasbile these days for non-booting volumes. I
just don't like to mix MD with LVM, that's my only concern.
----- Original Message ----
From: Ken Kauffman <kkauffman at headfog.com>
Anyone interesting in putting a real raid card in your systems, should
look into 3ware controllers. Be prepared to pay money though.
More information about the Novalug