Prev: hard-disk-drives Next: interlude-files-and-directories
RAID stands for Redundant Arrays of Inexpensive Disks. RAID allows a group of disks to act as one disk, with the benefit that using multiple disks in parallel improves its performance, and increases its redundancy RAID can tolerate the loss of a disk and operate as if nothing were wrong.
RAID also provides this advantage transparently to systems that use them. It looks just like a normal hard disk.
To a file system, RAID just looks like a big, reliable disk.
However, RAID is normally a separate hardware box with a standard connection (SCSI, SATA) to a host. Internally, they have firmware that synchronizes the disks and has DRAM for block buffering, as well as parity calculations for block writes.
One model to evaluate RAID is fail-stop, which means
that a disk is either working or failed. There are other failures like
disk corruption.
Three axes of evaluation:
The first approach isn’t redundant, so it’s not technically RAID. It
involves striping, or putting blocks in a round robin
fashion across the disks. Thus, reads and writes can progress in
parallel, and the capacity used is 100% of all blocks.
This might look like:
| Disk 0 | Disk 1 | Disk 2 | Disk 3 |
|---|---|---|---|
| 0 | 1 | 2 | 3 |
| 4 | 5 | 6 | 7 |
| 8 | 9 | 10 | 11 |
| 12 | 13 | 14 | 15 |
RAID blocks can be mapped like so:
Disk = A % number_of_disks
Offset = A / number_of_disksChunk size (the size of blocks before moving onto the next disk) affects performance. A small chunk size implies that files might get striped across many disks, which increases parallelism at the cost of positioning time.
To evaluate performance, there are two factors to care for:
As well as two broad workload types:
To calculate sequential and random disk performance, we could do the following tests on this disk:
Average seek time: 7 ms Average rotational delay: 3 ms Transfer rate of disk: 50 MB/s
To compute the sequential transfer rate, imagine transferring a 10 MB block:
The seek + rotational delay is the delay before the transfer starts (10 ms), and it takes 200 ms to transfer 10 MB at 50 MB/s. Thus, the whole transfer takes 210 ms, and the throughput is 47.62 MB/s.
For a random workload of 10 KB/s, the transfer takes 10 ms to rotate and seek, and takes 0.195 ms to transfer 10 KB at 50 MB/s. It thus takes 10.195 ms to transfer 10 KB, with a throughput of 0.981 MB/s.
RAID 1 involves mirroring each block to one other disk. This would look like:
| Disk 0 | Disk 1 | Disk 2 | Disk 3 |
|---|---|---|---|
| 0 | 0 | 1 | 1 |
| 2 | 2 | 3 | 3 |
| 4 | 4 | 5 | 5 |
| 6 | 6 | 7 | 7 |
When reading a block from a mirrored array, it can read from either copy. When writing, however, it must update both copies of the data. At least they can occur in parallel – a system that supports parallel writes would suffer less in performance.
From a capacity standpoint, RAID-1 is expensive. With mirroring for two copies, we can only use half our capacity. From a reliability perspective, RAID can handle anywhere from 1 failure to N/2 failures. Performance wise, reads are as fast as always, and writes either take about twice as long (with no parallelism) or about the same time (just a little worse, because of tail latencies).
Compared to RAID-0, sequential reads are the same performance, writes are half as fast, but random reads are just as fast as RAID-0, and random writes are half as fast.
With mirroring, writes issued to two different disks could enter a state where one fails and one succeeds. However, RAID wants a system that can write to disks atomically. The general way to solve this is to use a write-ahead log in non-volatile RAM. This way, consistent updating is provided without the high cost of logging to disk.
RAID-4 uses parity to add redundancy to a disk array.
They use less capacity than mirroring, at the cost of
performance. A Parity disk might look like this:
| Disk 0 | Disk 1 | Disk 2 | Disk 3 | Disk 4 |
|---|---|---|---|---|
| 0 | 1 | 2 | 3 | P0 |
| 4 | 5 | 6 | 7 | P1 |
| 8 | 9 | 10 | 11 | P2 |
| 12 | 13 | 14 | 15 | P3 |
Doing parity like this enables the loss of any single disk without data loss. This is done by XORing each bit that the parity disk holds. To recover, the parity bit and remaining disks are XOR’ed, which allows for reconstructing data on one disk.
RAID-4 is capacity efficient – it allows N - 1 disks to be used. For a sequential read, (N - 1) * S MB/s. For a sequential write, (N - 1) * S MB/s. Random reads are also (N - 1) * R MB/s. Random writes, however, are R/2 MB/s. This is due to having to update the parity value, so you must read at least the parity value and the previous value, and write the new parity value and the new value, so all writes cost 2 reads and 2 writes per block.
As above, RAID-4 has the problem of requiring updates to a parity value, which means that writes cannot succeed in parallel when updating parity corresponding to the same block. To fix this, RAID-5 has the parity strewn across the disks:
| Disk 0 | Disk 1 | Disk 2 | Disk 3 | Disk 4 |
|---|---|---|---|---|
| 0 | 1 | 2 | 3 | P0 |
| 4 | 5 | 6 | P1 | 7 |
| 8 | 9 | P2 | 10 | 11 |
| 12 | P3 | 13 | 14 | 15 |
| P4 | 16 | 17 | 18 | 19 |
This improves random read performance, and write performance, effectively making random writes progress in parallel, with a throughput of N / 4 R.
Prev: hard-disk-drives Next: interlude-files-and-directories