Concatenations are also know as "Simple" RAIDs. A Concatenation
is a collection of disks that are "welded" together. Data in a concatenation
is layed across the disks in a linear fashion from on disk to the next. So
if we've got 3 9G (gig) disks that are made into a Simple RAID, we'll
end up with a single 27G virtual disk (volume). When you write data to the
disk you'll write to the first disk, and you'll keep writing your data
to the first disk until it's full, then you'll start writing to the second
disk, and so on. All this is done by the Volume Manager, which is "keeper
of the RAID". Concatenation is the cornerstone of RAID.
Now, do you see the problem with this type of RAID? Because
we're writing data linearly across the disks, if we only have 7G of
data on our RAID we're only using the first disk! The 2 other disks
are just sitting there bored and useless. This sucks. We got the
big disk we wanted, but it's not any better than a normal disk drive
you can buy off the shelves in terms of performance. There has
got to be a better way..........
Striping is similar to Concatenation because it will turn
a bunch of little disks into a big single virtual disk (volume), but
the difference here is that when we write data we write it across ALL
the disks. So, when we need to read or write data we're moving really
really fast, in fact faster than any one disk could move. There
are 2 things to know about RAID-0, they are: stripe width, and columns.
They sound scary, but they're totally sweet, let me show you.
So, if we're going to read and write across multiple disks in our
RAID we need an organized way to go about it. First, we'll have to agree
on how much data should be written to a disk before moving to the next;
we call that our "stripe width". Then we'll need far kooler term for
each disk, a term that allows us to visualize our new RAID better.....
"column" sounds kool! Alright, so each disk is a "column" and the
amount of data we put on each "column" before moving to the next is
our "stripe width".
Let's solidify this. If we're building a RAID-0 with 4 columns,
and a stripe width of 128k, what do I have? It might look something like this:
Look good? So, when we start writing to our new RAID, we'll write the first 128k
to the first column, then the next 128k to the second column, then the next 128k
to the third column, then the next 128k to the fourth column, THEN the next 128k
to the first column, and keep going till all the data is written. See? If we
were writing a 1M file we'd wrap that one file around all 4 disks almost 3 times!
Can you see now where our speed up comes from? SCSI drives can write data
at about (depending on what type of drive and what type of SCSI) 20M/s. On
our Striped RAID we'd be writing at 80M/s! Kool huh!?
But, now we've got ANOTHER problem. In a Simple RAID if we had, say,
3 9G disks, we'd have 27G of data. Now, if I only wrote 9G of data to that
RAID and the third disk died, so what, there is no data on it. (See where
I'm going with this?) We'd only be using one of our three disks in a simple.
BUT, in a Striped RAID, we could write only 10M of data to the RAID, but
if even ONE disk failed, the whole thing would be trash because we wrote it
on ALL of the disks. So, how do we solve this one?
Mirroring isn't actually a "RAID" like the other forms, but
it's a critical component to RAID, so it was honored by being given
it's own number. The concept is to create a separate RAID (Simple
or RAID0) that is used to duplicate an existing RAID. So, it's
literally a mirror image of your RAID. This is done so that if
a disk crashes in your RAID the mirror will take over. If one
RAID crashes, then the other RAID takes its place. Simple, right?
There's not much to it. However, there is a new problem!
This is expensive... really expensive. Let's say you wanted a
27G RAID. So you bought 3 9G drives. In order to mirror it
you'll need to buy 3 more 9G drives. If you ever get depressed
you'll start thinking: "You know, I just shelled out $400 for
3 more drives, and I don't even get more usable space!". Well, in
this industry we all get depressed a lot so, they thought of another
kool idea for a RAID......
When we talk about mirroring (RAID-1) we're not
explicitly specifying whether we're mirroring a Simple RAID
or a Striped (RAID-0) RAID. RAID-0+1 is a term used to
explicitly say that we're mirroring a Striped RAID. The only
thing you need to know about it is this...
A mirror is nothing more that another RAID identical
to the RAID we're trying to protect. So when we build a mirror
we'll need the mirror to be the same type of RAID as the
original RAID. If the RAID we want to mirror is a Simple RAID,
our mirror then will be a Simple RAID. If we want to mirror
a Striped RAID, then we'll want another Striped RAID to mirror
the first. Right? So, if you say to me, we're building a
RAID-0+1, I know that we're going to mirror a Striped RAID,
and the mirror itself is going to be striped as well.
You'll see this term used more often than "RAID-1" simply
because a mirror, in and of itself, isn't useful. Again,
it's not really a "RAID" in the sense that we mean to use
the word.
RAID-5 is the ideal solution for maximizing disk space
and disk redundancy. It's like Striping (RAID-0) in the fact
that we have columns and stripe widths, but when we write data
two interesting things happen: the data is written to multiple disks
at the same time, and parity is written with the data.
Okey, let's break it down a bit. Let's say we build a RAID-5 out of 4 9G
drives. So we'll have 4 columns, and lets say our stripe width is 128k again.
The first 128k is written on disks one, two AND three. At the same time
it's written a little magic number is written on each disk with the data. That
magic number is called the parity. Then, the second 128k of data is written
to (watch carefully) disks two, three and four. Again, a parity number is
written with that data. The third 128k of data is written to disks three, four
and one. (See, we wrapped around). And data keeps being written like that.
Here's the beauty of it. Each piece of our data is on three different
disks in the RAID at the same time! Let's look back at our 4 disk raid. We're
working normally, writing along, and then SNAP! Disk 3 fails! Are we worried?
Not particularly. Because our data is being written to 3 disks per write instead
of just one, the RAID is smart enough to just get the data off the other 2 disks
it wrote to! Then, once we replace the bad disk with a new one, the RAID "floods"
all the data back onto the disk from the data on the other 2 adjacent disks!
But, you ask, how does the RAID know it's giving you the correct data? Because
of our parity. When the data was written to disk(s) that parity was written with
it. We (actually the computer does this automatically) just look at the data on
disks 2 and 4, then compare (XOR) the parity written with the data and if the parity
checks out, we know the data is good. Kool huh?
Now, as you might expect, this isn't perfect either. Why? Okey,
number 1, remember that parity that saves our butt and makes sure our
data is good? Well, as you might expect the systems CPU has to calculate that,
which isn't hard but we're still wasting CPU cycles for the RAID, which means
if the system is really loaded we may need to (eek!) wait. This is the "performance
hit" you'll hear people talk about. Also, we're writing to 3 disks at a time
for the SAME data, which means we're using up I/O bandwidth and not getting a
real boost out of it.
There are battles fought in the storage arena, much like
the old UNIX vs NT battles. We tend to fight over RAID0+1 vs RAID5.
The fact is that RAID5 is advantageous because we use less disks in
the endeavor to provide large amounts of disk space, while still having
protection. All that means is that RAID5 is inexpensive compared to
RAID0+1 where we'll need double the amount of disk we expect to use,
because we'll only need a third more disks rather than twice as many.
But, then RAID5 is also slower than RAID0+1 because of that damned parity.
If you really want speed, you'll need to bite the bullet and use RAID0+1
because even though you need more disks, you don't need to calculate
anything, you just dump the data to the disks. In my estimates (this isn't
scientific, just what I've noticed by experience) RAID0+1 is about
20%-30% faster than RAID5.
Now, in the real world, you rarely have much choice, and the
way to go is clear. If you're given 10 9G disks and are told to
create a 60G RAID, and you can't buy more disks, you'll need to either
go RAID5, or be unprotected. However, if you've got thoughs same
disks and they only want 36G RAID you can go RAID0+1, with the only
drawback that they won't have much room to grow. It's all up to you
as an admin, but always take growth into account. Look at what
you've got, downtime availability to grow when needed, budget, performance
needs, etc, etc, etc. Welcome to the world of capacity planning!