Todd Denniston wrote:
Roger Heflin wrote, On 03/24/2008 02:20 PM:
edwardspl@xxxxxxxxxx wrote:
Alan Cox wrote:
On Sun, 23 Mar 2008 15:39:22 +0800
edwardspl@xxxxxxxxxx wrote:
Dear All,
Which model / type of ATA Raid card controller is good for work
with New FC System ?
Would you please recommend ?
Almost every 'raid' controller for ATA devices is just driver level
raid,
so equivalent to using the built in lvm/md raid support that works with
any devices. At the high end there are a few hardware raid cards but
they
rarely outperform ordinary ATA on PCI Express.
Edward,
The cheapest 4-port raid cards are typically $300US, the 8-port cards
are quite a bit more. If you are a home user I would suggest not
wasting your money on the HW raid, and has others mentioned it is not
really worth the extra money for a home user, so use software raid.
Most of the cheaper cards are fakeraid and at best (if supported under
DMRAID) are only slightly better than software raid.
Roger
So would the better question be:
Which model / type of ATA multi-port card controller is good when you
want to do software RAID with New Fedora System?
i.e. which manufactures cards that you can hang 4+ drives off of, have
enough independence[1] between drives, that doing software RAID works
fast[2]?
Can you get 4+ port SATA cards that don't claim to be "RAID" cards?
Yes. The problem is if you have a PCI (not -E or -X) the bus is limited to
about 130 MB/second for all cards, so for a fast machine you need either -E or
-X and multi-port cards of those get expensive too.
I have a crappy (but cheap) 4-port SATA SIL card, it works, it is PCI and not
the fastest but it is cheap, and appears to be reliable, it only does about
60MB/second writes and 90MB/second reads with 4 disks under RAID5.
If you need speed and more than 4 ports, the cost goes up quite a bit.
And you have to test them, I have seen cards that each 2 sata ports share
hardware, so using 2 disks on ports 1 and 3 is faster than using 2 disks on
ports 1 and 2, and of course all of the ports share the bus the card is plugged
into, so it is critical to test things as none of the specifications will
actually tell you any of this.
A lot of the highpoint fakeraid cards are based on the Marvell sata chipset(s)
and perform pretty good, but you have to actually confirm that the given version
is truly supported under Linux without their extra driver, and they are
reasonably cheap. I have not tested the recent multiport Adaptec controllers
so I don't know what they will do.
On motherboards you have a similar set of issues, depending on where the given
motherboard connects in the Sata controllers depends on how much the total
throughput could be. Some put the sata controllers on the PCI (non-X non-E)
bus part which makes them truly suck, and some put things closer to the cpu and
on higher bandwidth parts of the mb/chipset, you need to look at the MB's block
diagram to figure out exactly what is shared with what to determine the best
config to get the most speed out of things, and what other components on the MB
can affect you (network..)
Or has everything already been said here:
http://linuxmafia.com/faq/Hardware/sata.html
http://linux-ata.org/faq-sata-raid.html
[1] I am making the old assumption that ATA drives on the same bus slow
each other down. Does that really matter with SATA?
Depends on the SATA card's chipset, generally with the newer ones it does not
matter.
[2] assuming the controller card is more likely to be the bottleneck
than the processor, PCI bus, or drives.
Unless you have a huge number of disks, and good bandwidth everywhere, the
answer is one of the busses to the SATA card will probably be the bottleneck, I
have had 3+ year old quad socket MB's with PCI-X sustain 360MB/second read or
writes with the bottleneck in that case being the 2-2Gbps (about 440MB/second
limit) fiber channel ports being used, in that case though I had several
external raid units attached to the machines each of which was quite fast, and
had I used more ports on separate PCI-X buses could have probably easily
exceeded that rate, but the spec we were trying to meet was met by the
360MB/second so there was no need to go to more extreme measures.
It comes down to, you need to know how a given MB allocates its bandwidth, and
what the limits are where things join together to even be able to guess before
testing what is going to be the limit, and you need to know what you actually
need for a given application.
Roger
