To Raid 0 or not too Raid 0

Controller overhead. You're increasing the complexity of the system therefore if the controller is not sufficiant to handle it the array loses speed.

Theoretical structures often don't bear out completely in an actual implementation.
 
ashmedai said:
Controller overhead. You're increasing the complexity of the system therefore if the controller is not sufficiant to handle it the array loses speed.

Sufficient in what regard? Computing power? All of the operations involved are trivial. And if it were computational overhead, why wouldn't the issue still manifest itself during longer transfers?

RAID 0, remember, doesn't involve checksum or parity computation. It only involves multiplexing logical requests to two physically different servicing objects.

The complexity increase is between nil and trivial. The operations required can be done in only a few clock cycles in an FPGA. As far as I can tell, they're required only initially or once per stripe-size transfer.

Am I wrong about that? Which ones aren't? What significant complexity has been added? What operations turn out to be very expensive, particularly compared to the overhead of the I/O operations themselves? Why can't those expensive operations be completed concurrently during the very large time the processing power (or logic) on the controller card is idle -- during the transfer?

I'm sorry to be so obstinate, but I'm loging for a detailed explanation. And a verifiable one? That would -- oooh. That would just make me so very, very happy.
 
I think it's closer to that you get some overhead regardless, but decreasingly so the better the controller is.

To elaborate on the cause of this:

RAID 0 tries to improve certain aspects of drive performance by striping the volume across two seperate physical drives. This helps with some things because if the file is sufficiantly large it will take up space on both drives, therefore each drive only has to do half the work it normally would, and you see nearly double the potential sequential transfer rate.

Seek time is frequently a smidge slower because instead of having to find data on a hard drive, it's finding data on an array that is made up of hard drives. Extra layer, more work, slows it down a bit no matter how good your controller is but increasingly so for cheap implementations.

For small files - they're not big enough to benefit from the striping and are just stored on one drive. Therefore you get the overhead penalty but no benefit from it being stored on a RAID 0.

That help any or am I forgetting to address anything?
 
ashmedai said:
That help any or am I forgetting to address anything?

I'm sorry, but there are still holes. It's nothing personal. I'm asking about this because I want to gain very deep and sound knowledge. If someone at NASA asks why we can't fly to the moon, "because it's pretty far away" is not enough detail.

You've fallen short of explaining what the "extra layer" is exactly, and what the overhead really is. Exactly: I'm sorry, but I'm not settling for the "it's just overhead" answer simply because I want to understand the problem and the implementation very deeply.

You get some overhead regardless of what, exactly? Transfer size? RAID 0 or not?

Seeking is not "finding" data. It's moving a physical device to a known physical location:

Finding: "Read track 0. Did that contain the key we wanted? No. Read track 1..."
Seeking: "Position the heads over track 373, please."

The time to seek is limited by the drive. A decent drive these days takes about 8ms to seek. A great drive these days takes about 3ms. In 3ms, a Pentium 90 (not even a Pentium Pro!) can do about 21,600 floating point division operations. (125 cycles per divide, 90,000,0000 cycles per second, 0.03 seconds.)

Of course, the results of all 21,600 operations have to be checked for accuracy because of the FDIV bug, but I think you get my point: If two commands are necessary, one command can be issued. While we're waiting for the pyhsical activity that command causes to actually complete, we have the opportunity to do an assload of other work before we have to answer the phone again.

Why are you paying overhead for something you don't use? If the file is so small that it only requires a read from one drive, why is there any activity at all, period, on the second drive? If the seek-then-read operation happens on only one drive, why is it any slower than using only one drive alone?

See the algorithm I suggest in this thread. Perhaps that thread is also a better place to continue the discussion, anyway.
 
I'm a college student, not an engineer, so if I can even approach the level of detail you're after I feel good about it. :D
 
Back
Top