Looking to build storage of about 3TB -- recommendations?

Jim

The appllications to run on the Opteron chips will actually be all
loaded into memory before run. Tun times are to be about 6hrs so the
IO will be a very small portion of that.

We are doing HW based RAID with 3Ware controllers.

daben

In article .com,
"daben" wrote:

Hi All

For my research I need large storage for satellite data files. I am
looking to build storage for about 1TB redundant array and 2TB
volatile array. I have looked around at snap servers or other SAN
solutions. I also have looked at scsi solutions but don't think that
scsi is economicable. What I am proposing is to build 2 arrays of SATA
drives with the following setup:

1 3u supermicro chasis with 15 hotswappable sata bays
2 3ware RAID controller (for ex 9550SX-8LP)
12 WD SATA HDS (400GB)

RAID ARRAY 1: RAID 5, 5 drives total, 1 hot spare == 1.2TB
RAID ARRAY 2: RAID 0 (or none), 7 drives total == 2.8TB

This will all be hooked to dual xeon chips with and internal boot/sys
mount. Will probably use win2k3 server due to software requirements.

The second array is for temporary storage and can crash and we will not
loose much except the need to download again. The first array needs to
be fault tolerant. We envision swapping out drives in array 2 as
larger drives become available.

My questions a

+ Can anyone propose reasons NOT to do this?

I have seen, for whatever odd reason, a high rate of failure of the SATA
cables. I have also seen backplane failure in these types of boxes.
Ironically, as some will see it, the SATA drives themselves seem to be
quite reliable.

+ Are the better solutions?

A self-contained controller is probably simpler and presumably less
likely to fail.

+ NOTE: We anticipate this will cost $5k-$6k -- the other solutins are
MUCH more that I have seen

Promise VTrak is not expensive and actually seems to be reliable. It
also has decent management software.

(Someone else mentioned RAID 6; that's the one thing the VTrak lacks.)

In article .com,
wrote:

Be a little careful here on the Opteron vs. Xeon speed comparisons. In
raw compute speeds the Opteron is definately faster. But the Intel
chipsets support a much higher performance in I/O throughput.

Do you by chance have a reference for this?

flux writes:

Promise VTrak is not expensive and actually seems to be reliable. It
also has decent management software.

The experience I've had with Promise cards is that they're mostly
software RAID (a lot happens in the driver). This is under Linux.

I'm not familiar with the VTrak specifically though.

I've heard many good things about 3ware.

--
David Magda dmagda at ee.ryerson.ca
Because the innovator has for enemies all those who have done well under
the old conditions, and lukewarm defenders in those who may do well
under the new. -- Niccolo Machiavelli, _The Prince_, Chapter VI

In article ,
David Magda wrote:

flux writes:

Promise VTrak is not expensive and actually seems to be reliable. It
also has decent management software.

The experience I've had with Promise cards is that they're mostly
software RAID (a lot happens in the driver). This is under Linux.

The VTrak is a completely self-contained hardware RAID.

I'm not familiar with the VTrak specifically though.

I've heard many good things about 3ware.

Ironically, 3ware may be problematic under Linux. I've occasionally
experienced "Out of IOMMU" space errors and such an error generally
spits junk data on the RAID filesystems, which is not good. I don't
know, however, if the source of the problem is the 3ware driver (is it
really "hardware" RAID?), the Linux kernel, the card itself, or
something interaction with the motherboard (Arima HDAMA) or the (Wintec)
memory. But since I don't see this problem on the LSI/mptfusion
card/driver in the same box, I'm inclined to believe it's the 3ware
driver or the card itself.

This is what I do for a living. But this particular analysis is pretty
easy to come up with.

Just look at the block diagram of the Intel E7520 chipset and then one
for the AMD 8131. What you are looking for are the number of
independent PCI buses and their paths to memory.

Each PCI bus has a maximum burst throughout, and a sustainable
throughput of about half of that (depending on the number of bus
masters). As long as the memory controller is not the bottleneck, then
it is the number of PCI buses times the sustainable throughput.

-Jim