Reliability, Availability and RAID

• Some definitions:
• Reliability
• Refers to the dependability of individual components of a system.
• Availability
• Is the system still available to the user after a failure of one or more of its components ?
• Adding hardware can therefore improve availability but it can NOT improve reliability .

 

• Disk arrays :
• The basic idea behind disk arrays is that by adding disks and therefore more disk arms working in parallel, bandwidth is improved.
• Individual process-request seek latencies can be overlapped in time.

 

• This is cost effective since price/megabyte is independent of disk size.

 

• However, latency for small requests is not improved because it still takes all of the usual latency to get to a randomly selected block.

Reliability, Availability and RAID

• Striping
• In disk arrays, the data from files can be striped across several disks.

 

• This increases bandwidth by allowing a file to be read from more than one disk at a time.

 

• The data is distributed round robin between the disks.

 

 

• Problems with disk arrays:
• Reliability
• Disks have a mean time to failure of about 20 years.

 

• However, a collection of 8 disks will experience a failure (on average) every 2.5 years.
• N devices generally have 1/N the reliability of a single device.

Reliability, Availability and RAID

• Problems with disk arrays:
• Reliability
• Increase the number to 1000 disks, and a failure occurs every 1/50th of a year or every week !

 

• When a disk fails, it takes its data with it.

 

• Availability
• The other problem with disk arrays is that the disk array becomes unusable after a single failure.

 

• We need a scheme to prevent data loss when a disk fails, and to allow the system to recover from the failure (remain available.)

Reliability, Availability and RAID

• The solution: Redundant Arrays of Inexpensive Disks.

 

• Improves availability by adding redundant disks.
• When a disk fails, the lost information can be reconstructed from redundant information.

 

• This works since the mean time to failure (MTTF) of disks is long (years) and the mean time to repair (MTTR) is short (hours).

 

• The idea behind RAID is to use some of the disks to store error correction information for the rest of the disks.

 

• RAIDs do NOT have to detect errors.
• The ECC kept on each sector by each disk allows the disk electronics to check and detect disk failures.

RAID Levels

• There are 7 levels of RAID, each of which can be characterized by their availability and overhead.

Raid Level		Failures survived	Data Disks	Check Disks
0	Nonredundant	0	8	0
1	Mirrored	1	8	8
2	Memory-style ECC	1	8	4
3	Bit-interleaved parity	1	8	1
4	Block-interleaved parity	1	8	1
5	Block-interleaved distributed parity	1	8	1
6	P+Q redundancy	2	8	2

• RAID 0
• This is no redundancy at all.
• However, it is the fastest and cheapest RAID level.

 

• This refers to the nonredundant disk array discussed previously.

RAID Levels

• RAID 1
• Disks in this configuration are mirrored or copied to another disk.

 

• With this arrangement, the data on a failed disk can be easily replaced by reading it from the other disk.

 

• In addition, reading is actually faster than it is for RAID 0 because read requests to the same disk can be split between the two disks.

 

• Writing is a little slower , though, because the file system must wait for the slower of the two requests.
• Remember, both disks must be updated.
• Also, seek time is different between the two disks since they are not synchronized.

 

• The main problem with RAID 1 is that it imposes a 50% space penalty.

 

• Therefore, it is the most expensive solution.

RAID Levels

• RAID 3 (Bit interleaved parity)
• In this scheme, data is striped across disks in very small units .

 

• These units are so small that all disks must work together on both reads and writes because a single sector actually spans all of the disks.

 

• Redundancy is implemented by calculating parity and storing it on the check disk.

 

• The overhead for n data disks is 1 disk, for a storage efficiency of n/(n+1) .

 

• This level can survive a single disk failure and reconstruct data using the remaining disks.

 

• This RAID level is somewhat limited since the entire disk system can only handle one request at a time.
• Thus, the sustainable request rate is no higher than that of a single disk.

RAID Levels

• RAID 4 (Block interleaved parity)
• This is similar to RAID 3, but individual disks each have a block (or more) of consecutive data within a stripe.

 

• This means that each disk can handle an individual small read request if all disks are working.

 

• However, writes are still a problem since the parity disk must participate in all write operations, which creates a bottleneck.

RAID Levels

• RAID 5 (Rotated parity)
• The biggest problem with RAID 4 is that the parity disk is a bottleneck.

 

• In RAID 5, the parity information may be stored on any disk.

• Advantages:
• Reads can be spread among all n+1 disks.
• For writes, the parity disk bottleneck is eliminated.

RAID Levels

• RAID 5 (Rotated parity)
• With respect to the parity calculation and update, RAID 5 does have a problem writing small pieces of data.

 

• If a full stripe is written, the parity calculation is simple to implement.
• XOR all the data together to get the parity.

 

• If only a single block is written, though, the system must somehow figure out the new parity.

 

• The most straightforward solution is to read the rest of the blocks in the stripe and XOR them together.

 

• A better solution involves only four accesses:
• Read the old data.
• Compare old data to new data (to determine which bits change).
• Read and change the old parity.
• Write the new data and new parity.

RAID Levels

• RAID 6 (P+Q parity)
• The objective of RAID 6 is to survive two disk failures.

 

• This is done using schemes such that only n out of n+2 disks are necessary to reconstruct the data.

 

• This RAID level is not common today because disk failures are still relatively rare.

 

 

• If a disk in a parity-based RAID fails, no data is lost if a new disk is installed and "updated" before another crash takes place.

 

• Since stripes tend to be relatively small (usually less than 32 disks), the chance that another disk will fail is relatively low.
• Most people don't worry about it.

RAID Issues:

• Mapping data to disks.
• The first problem in a RAID system is how to map data to disks.
• Usually, this is done using simple modulo arithmetic.

 

• It is more complex for RAID 5, but there is still a formula that allows you to determine where a given block is located.

 

• Reconstruction
• After a disk has failed, it is replaced.

 

• But we are not done since the new disk must be reconstructed with the lost information in order to allow the array to withstand another crash.

 

• Since the lost information is "embedded" in the remaining disks, reconstruction is possible and carried out immediately.