Monday, June 04, 2012

Guide to RAID

There are many different levels of RAID, what I intend to do here is explain the most popular levels of RAID. which I believe to be RAID 1, RAID 5 and RAID 6. I will also discus RAID 0, and exactly why this is not actually a RAID level.
as it is useful, I will also explain the concepts of RAID 1+0.

RAID means Redundant Array of Independent Disks

Basically, using RAID is a way to mitigate the effects of a a hardware failure.

In each case what using RAID does, is extend data across multiple disks, with a view that should one disk fail, you'd be able to recover the data from the remaining disks without going through the trouble of sending the drives off to a data recovery company.

RAID 0, (striped disks)
In RAID 0 you must have at least 2 drives, however you can feasibly have any number of drives up to the limit that your controller will support. Ideally your drives must be of the same size, though that is not a definite requirement, however, if your drives are not the same size you will loose some of the capacity of the bigger drive.

the first term that you will need to understand is a stripe, a stripe is simply a block of data on a disk.

01) if you image that
02) this sentence is 
03) a piece of data, 
04) then you can see 
05) how it has been d
06) ivided into 10 eq
07) ual strips (repre
08) sented by a new l
09) ine) each 17 char
10) ectors in length.

so in the above example the block size on the disk is 17 bytes. so that's the size of a stripe.

With RAID 0 each one of these stripes is written to a separate disk
imagine a two disk example disk 1 on the left, disk 2 on the right.

01) if you image that  this sentence is 
02) a piece of data,   then you can see 
03) how it has been d  ivided into 10 eq
04) ual strips (repre  sented by a new l
05) ine) each 17 char  ectors in length.
you can also see that where as when all the data was written on one disk it filled the disk (which was only ten lines big), by spreading the data across 2 disks, there is an effective doubling of capacity.

This is why disks must be equal size, is disk 1 can only contain 10 lines, but disk 2 is twice as big, then the RAID can only use as much space as it can equally allocate between the disks.

you can also see that if one disk dies we'll be left with this:
01) if you image that   
02) a piece of data,    
03) how it has been d  
04) ual strips (repre  
05) ine) each 17 char  
basically, lots of corrupt files,

RAID means redundant array of independent disks, whist the disks in RAID 0 are an array of independent disks, there is no redundant hardware.

Incidentally if you would like to make 1 large volume from two different sized disks then you should investigate extending the volume.

What are the pro's of RAID0?

Makes bigger volumes
Faster read and write times as data can be read or written from two independent disks, (almost doubling the read/write time with two disks, tripling with three disks).
you can have any number of disks.

What are the cons of RAID 0?
It's not redundant, loosing 1 disk means loosing all the data
The more disks that you have, the more likely it is that your array will die sooner rather than later.

Probability of failure.
whilst of course there may be somewhere in the world a RAID 0array with fifty disks that has never failed, and yet plenty of 2 disk arrays that have failed.

Failure is never assured, but neither is it assured that the disk will keep running.

think of it this way, if the likelihood of a disk failing is 1 in a million. when you have two disks the likelihood of either one failing is 2 in one million.

if you have 20 disks, it's ten time more likely that one disk will fail and you'll loose all the data than if you had two disks. in a RAID 0 array a single failure means a complete loss of data, for this reason it's generally not advisable to use RAID 0, unless you really understand that you're likely to lose data, but you really require the increases in performance that this RAID level offers.

RAID 1 (mirrored set)
in RAID 1 you have to have to have at least 2 disks, however you can have three or more disks. with RAID 1 there is no increase in usable space of capacity. there is no striping, data is not divided into stripes. data is mirrored at a disk level.

again a visual example, disk 1 on the left, disk 2 on the right
01) data goes here  01) data goes here
02) more goes here  02) more goes here
03) even more data  03) even more data
you can see that whatever is written to disk 1 is also written to disk 2

if there are three disks the same data is written to all three disks
01) data goes here  01) data goes here  01) data goes here
02) more goes here  02) more goes here  02) more goes here
03) even more data  03) even more data  03) even more data
if one disk fails, then you can see that you can still read data from another disk, if you have 3 disks in a RAID 1 array then 2 disks could fail and you would still be able to access your data.

Fast access to data, you can read simultaneously from as many disks as you have in your array, if you have 5 disks, you can read data five times as quickly.
it is fault tolerant, a disk can fail, and you won't loose any data.

waste of disks, if you've got an array of two disks, you're effectively wasting the space of one of them, if you've got three disks, you're wasting two of them. -though at the benefit of two disks being able to fail at once.

RAID 1+0 (mirrored striped)
RAID 1+0 expands on the concepts above. you must have at least four disks to create a RAID 1+0 array.

Basically what this will do is create 1 set of 2 disks where data is striped across the two disks, and then also mirror that data to a second set of stripped disks.

Moving on from here, you're going to need to have that data stripe theory right in your mind.
you'll also need to be able to grasp what parity is.

Parity is something really clever, that can be explained really simply.
you can see that the problem with just mirroring (RAID 1) a piece of data is that it's incredibly inefficient. it also is restrictive in terms of volume size. if you've got 2 100GB disks, in order to have a 200GB array you have to throw away both 100GB disks and add 2 200GB disks.

How Parity Works:
I'm going to explain bit level parity using the example of 3 bits, then I'll expand to larger bit sizes so you can see how well it scales.

with parity, you always loose a bit, or in this example of RAID arrays a disk to parity.
a parity bit is a bit that would enable you to work out what you're missing if a disk fails in this example there are 2 disks for data, and one disk (or bit) for parity
1 2 P
0 0
0 1
1 0
1 1
if you look above you can see that the disk has a series of data written to it. in order to calculate the parity, all you do is add up the amount of 1's that are written to the disk, if it's an odd amount you write 0, if it's an even amount you write 1, (and for the sake of parity, 0 is an even number)
1 2 P
0 0 0 (there are no 1's, 0 is an even number so write 0 for parity)
0 1 1 (there is one 1, it's an odd number so the parity bit is 1)
1 0 1 (one 1 means an odd number of 1's, so the parity bit is 1)
1 1 0 (2 ones, 2 is an even number so the parity bit is 0)
no imaging that disk 1 fails
1 2 P
? 0 0
? 1 1
? 0 1
? 1 0
thankfully, because you have a parity bit you can calculate what it's meant to be
1 2 P
? 0 0 (well there was an even number of ones when the data was written, and disk 2 is a 0, so disk 1 can't be a one, it must be a 0)
? 1 1 (parity bit = 1, so there must have been an odd number of 1's, disk 2 is a 1, so disk 2 must have been a 0)
? 0 1 (parity bit = 1, so there must have been an odd number of 1's, disk 2 is a 0, so disk 1 must have been a 1)
? 1 0 (parity bit = 0, so there must have been an even number of 1's, disk 2 is a 1, so disk 1 must have also been a 1)
whilst the disk has failed, you can still see what was on the disk by virtue of the fact that you can calculate it, you can put a new disk into this array, and the data that was on the disk can be recalculated from the parity bit, and re-written to the new disk, as if nothing ever happened whilst it may not be immediately clear, parity can work on any number of disks
1 2 3 4 5 6 7 8 P
1 1 1 1 0 1 0 1 0 (there are six ones, it's an even number so the parity bit is 0)

1 2 3 4 5 6 7 8 P
1 1 ? 1 0 1 0 1 0 as the parity bit is a 0, you know that there were an even number of ones, disk 3 has failed leaving an odd number, so now you know that there must have been a 1 on that disk.

1 2 3 4 5 6 7 8 P
1 1 1 1 0 1 ? 1 0 as the parity bit is a 0, you know that there was an even number of ones, disk 7 failed, but since you know that there are 6 1's in the remaining disks, and there were an even number of 1's that the data on disk 7 must have been a zero.

from this you can see that as you add disks to the array, you increase the size of the array, not only that but a disk (any disk) can fail, (1 at a time!) and your data is still safe.

whilst I didn't say at the start, any number of disks, but with a dedicated disk reserved for recording parity information is RAID 3, or RAID 4.

RAID 3 works at byte level, RAID 4 works at disk block level.

the exact parity method isn't as described above, -that's just a really simple example to make it easy to understand that it's possible to use just a single disk to protect an array of many other disks

    1 2 3 4 5 6 7 8 P
01) 1 1 1 1 0 1 0 1 0
02) 1 0 1 0 0 0 1 0 1
03) 1 0 0 0 0 1 0 0 0

RAID 5 (Distributed Parity)
RAID 5 works almost exactly like RAID 4, you must have at least 3 disks, and will always loose the space of 1 disk to parity.

if you have 3 100GB disks you'd have 200GB of space, (200GB of data 100GB of parity information)
if you had 10 100GB disks you'd have 900GB of space (900GB of data 100GB of parity)

basically, RAID 5 differs from RAID 3 in that instead of having a disk dedicated to Parity, the parity block is spread around the disks,

where D = data and P = parity, the disks will look like this
    1 2 3
01) D D P
02) D P D
03) P D D
04) D D P
RAID 3, RAID 4 and RAID 5 all have increased read rates as data is read from multiple disks, both suffer from reduce write rates, due to the extra time taken to process the parity bit.

RAID 5, like RAID 4 will allow for any one disk in the array to fail.

RAID 6, is a lot like raid 5 in that there is parity, and the parity blocks are spread across all the disks, however, RAID 6 uses proprietary methods for calculating the partity, and stored parity information on 2 disks in the array.

this means that an array of 15 100GB disks would not have 1500GB space it'd actually only have 1300GB of space that'd be 200GB of parity information.

For this reason to get the best protection and efficiency RAID 6 is best used with large arrays, (12 disks or greater), and the proprietor calculations mean that hardware acceleration is almost essential.

It is possible, either through the use of specialist hardware controllers, that support this feature, or a mix of hardware controllers and software configuration. to effectively multiple RAID levels.

it was mentioned earlier that RAID 1+0 was a possible configuration, through really requires 4 disks, and at least 2 extra disks added each time the array needs to be expanded.

The same is true of other RAID levels, you can have the protection of a Level, and add a second layer of protection on top of that. (which is better for very large arrays).

for example there may be twenty disks, each disk might be 1TB

a RAID 0 array is very attractive, 20TB of space in total, but it's twenty times more likely to fail than if there was a single really big disk, (even though such a disk doesn't really exist). Basically, a RAID 0 array is a bad choice, it's very likely to fail.

a RAID 1 array is a terrible choice, you'd get a 1TB disk, mirrored to 19 other disks! a massive waste of hardware!

A RAID 4, or 5 array is a possibility, at least at this point you're going to have 19 of your 20 1TB drives essentially added together, but with 20 disks in your array, it's likely that there may be multiple failures.

so really this leaves 3 possible options...

You use RAID 5, but you allocate one drive as a hot swap drive, this means that the drive sits unused until another drive fails and it's needed
1  2  3  4  5  6  7  8
D1 D2 D3 D4 D5 D6 D7 HS
This is a RAID 5 array.

how Imagine that drive 3 fails, the Hot spare takes over, and the array is rebuilt onto that drive. then when drive 3 is replaced, either the array will rebuild onto that drive leaving disk 8 as the hot spare again, or drive 3 will become the hot spare until such a time as another disk fails. then it'd assume that responsibility.
1  2  3  4  5  6  7  8
D1 D2 XX D4 D5 D6 D7 D3
in the case of this example, clearly you'd have 20 1TB disks, 1 is taken away for hot spare, 1 is taken away for Parity, you'd have an 18TB array.

either the hot spare, or one of the disks in the array could fail. at the same time(!)

You could also use use RAID 6,
this will lead to you having an 18TB array (18 data drives, 2 parity disks), you have better fault resistance because there are 2 Parity disks but there will be an increase in the time taken to write data to the disks. due to the fact that the parity has to be calculated using a complicated mechanism at each data write.

Lastly, you could have raid 5+0.

What this means is that you'd have 2 arrays, each would have 10 disks, arranged in a RAID 5 array, (9 data disks, 1 Parity disk) each array would be 9TB
then you add these disks together using a RAID 0 array.
This would give you a fairly well protected array, (any one disk from either array could fail) -you could have 2 disks failed at once, just one from either array...
you also get the added benefit of the increased read speed. (would be the same speed as RAID 6 as data is being read from 18 individual spindles), but the write operations would be faster as the parity calculation is easier, (unless you have hardware acceleration for RAID 6 calculations).

RAID 2 is intentionally omitted because practically nobody uses it.

(you're about halfway through!)

I'll say one thing, that's that whilst it's important to understand that RAID offers levels of protection, it's not critically important to understand how it all works, just what kind of level of protection you have, and how that effects solutions that you put in place.

RAID doesn't (usually) really relate to storage solutions for home, these are more the kinds of things that you'll be implementing in a business. certainly he last part about RAIDed RAID is really only the kind of stuff you'd be doing at a very large SAN type level, where you'd have a single controller and then perhaps 8 "shelves" which might contain 12 or 14 disks each...

anyway... where you might use these.

RAID 0 (striped) is a home solution, you're going to use this when you have perhaps 2 small drives, and you'd like to have one big drive, for example you have a couple of 250GB drives, but you really want a single volume of 500GB.
you'll use RAID 1 to add these together.

the fact that if one drive fails then the whole array fails, (and the fact that as you add more disks it becomes more likely that at least one would fail), means that RAID 0 is really not suited to business applications at all (unless you're combining it with other RAID solutions.

RAID 1 (mirrored) is the kind of thing you'll only be using on a business critical server.
usually in this example you'll be using hardware RAID to have the OS installed across 2 disks, of say an important domain controller at a specific location.

this means that if one disk in the array fails your machine will keep working whilst you're organising a replacement, with the best will and good hopes in the world, all disks ARE going to fail one day, the point is that you can't guarantee that you'll be able to replace that disk the same day.
in the case of a domain controller you might be talking about an office of 400 people that can't log on in the morning, if you have to send them home for the day because you can't replace the disk till the next day then you've lost over a man year of time (which is expensive!).

basically, you've got a disk that wasted, but when the time comes that one disk in the array fails, you'll be glad that you weren't relying on just that one disk, the cost of a man year of time, (thousands) is going to far outweigh the extra 100 spent at the start!

RAID 3 and RAID 4 (striped with dedicated parity) pretty much aren't really used anywhere in favour of using Raid 5 (striped with distributed parity).

you're going to use RAID 5 where you have at least three disks, and you want to protect those disks from failure.
an obvious (but contentious) example is when you're talking about database servers.

RAID 5 is well suited to database servers.

the read speed is very good (due to the use of multiple spindles), the write speed suffers slightly (so you may wish to consider either hardware acceleration or a different RAID level if you have a database that is very write intensive, or has frequent random data writes).

In addition to the faster read speed (than from a single large disk). you also have the expandability.

say for example you have a SCSI enclosure, you could start your enclosure with just 3 150GB drives, (giving you 300GB of space). and protection from the parity drive.
after a couple of years you may start to run out of space, you you can buy another 150GB drive. now you have 3 data drives and one parity (450GB of space).

a year later you add another drive and get 600GB of space for data.

Basically, the use of RAID 5 enables you to build your device as a starter device which is still protected. but is cheap because you start with a relativity expensive enclosure and 3 relativity inexpensive disks, over the course of time you can increase your storage space as required with a small investment, just the cost of the disk).

RAID 5 tends to only be used in business, (there aren't a lot of software implementations so specialist RAID cards that support the RAID level are required).

RAID 6 would again be used when you're setting up a drive that requires plenty of space, that needs to be protected but quite cheaply. (only loosing 2 disks to parity rather than loosing half the disks as you would in a mirrored scenario).
the difference between RAID 5 and RAID 6 (aside from the method of parity and amount of parity) relates to the amount of disks.

basically, you're going to use RAID 5 if you have say 10 disks or less.
RAID 6 if you have 12 disks of more. (11 disks, you may choose to use either!).

The reason for this is :
Probability of failure.

Basically it works like this,

there is a mean time to failure, this is expressed as the likelihood of failure for a given time,

say you have a batch of 100 disks, you might expect that over the course of ten years.
1 disk fails in year 1, (99 live on)
2 disks fail in year 2 (97 live on)
5 disks fail in year 3, (92 live on)
10 disks fail in year 4, (82 live on)
20 disks fail in year 5 (62 live on)
30 disks fail in year 6 (32 live on)
20 disks fail in year 7 (12 live on)
10 disks fail in year 8 (2 live on)
2 disks fail in year 9 (no disks from this batch or model live past 9 years) That's actual failure not just replaced and put in the bin.
(if you draw that it should make a reasonably nice Gaussian distribution)

what this means is that the probability of a failure of a disk in year 1 is 1 in 100 (0.01)
probability of a disk failure by year 2 is 3 in 100 (0.03) (3 out of 100 will have failed by year 2)probability of a disk failure by year 3 is 8 in 100 (0.08)
probability of a disk failure by year 3 is 18 in 100 (0.18)
probability of a disk failure by year 4 is 38 in 100 (0.38) (basically a 1 in three chance that you'll wake up to a failed disk)
probability of a disk failure by year 5 is 68 in 100 (0.68)
probability of a disk failure by year 6 is 88 in 100 (0.88)
probability of a disk failure by year 7 is 89 in 100 (0.98)
probability of a disk failure by year 8 is 100 in 100 (1)
probability of a disk failure by year 9 100 in 100 (1)

Imagine that you were using RAID 0 (striped with no mirrors or no parity) failure of 1 disk means you loose all your data.

what you're saying is that in the first year there's a 1 in 100 chance that all your data will go.

if you're using 100 drives in your RAID 0 array, that's basically saying you're going to loose all your data in year 1. (because statistically it's a certainty that 1 of those 100 will fail). -of course in life the statistics might not play out exactly.

if you're using RAID 5, or RAID 6 it's like saying that 1 drive will fail, but that's ok, you can replace that.

Fast forward to year 6, so far, you'll have replaced 38 out of your original order of 100 drives.
but now in year 6, it's expected that 20 drives are going to fail. if you're using RAID 5 then this becomes a problem.

if the drive failures were spread evenly throughout the year, you're expecting a drive failure roughly every 17 days.

and inevitably it'll play out like this (we're talking a big business here).

day 1 drive fails
day 2 you log a ticket to the help desk telling them to tell the procurement team to buy a new disk
day 3 the helpdesk log the ticket
day 5 the procurement team ask for a quote
weekend happens
day 9 the procurement team get a quote back from the supplier.
day 10 this quote is sent to a manager for approval
day 12 the quote is approved
another weekend happens
day 15 the procurement team ask for the disk to be ordered
day 16 the supplier receives the order and asks for another disk to be sent.
day 17 The disk is posted to you
day 17 another disk fails and all your data is lost!
day 19 the disk arrives with you perhaps even at the start of the day -it doesn't matter anyway, at this point so many disks are failing that you can't replace them as fast as they are failing.

Using RAID 6 means that 2 drives can fail at once. which effectively give you twice as long to replace the drives, in the example above, it doesn't matter that a second disk failed halfway through the procurement process, because in RAID 6 two disks CAN fail at once.

the above example is very simplistic and assumes that drive will fail predictably, this is more or less accurate, but does ignore the fact that specific manufactured batches of disks may be pre-disposed to failure.

you might end up with a bad batch and have all drives fail in year 1!

which brings us nicely to RAID + RAID.
Basically, you're only going to want to use RAID + RAID when you have many (many disks), and there isn't another way around it.

here I'm talking about very large SAN implementations, where your RAID controller is a physically server sized device, connected to many rack mounted shelves containing bunches of disks, (12 or 14 usually for individual shelves).
AND you want those drives to be presented as en enormous single data volume.

Practically, if you're designing solutions to that level then you're going to need a better understanding than I could give in a short guide.

(basically because at that level you're running into the risk and probability factors not only of disks, but of disk enclosures, interconnecting cables, head units, power strips, backup generators etc...)

No comments: