1 |
On Sun, May 3, 2020 at 5:32 PM antlists <antlists@××××××××××××.uk> wrote: |
2 |
> |
3 |
> On 03/05/2020 21:07, Rich Freeman wrote: |
4 |
> > I don't think you should focus so much on whether read=write in your |
5 |
> > RAID. I'd focus more on whether read and write both meet your |
6 |
> > requirements. |
7 |
> |
8 |
> If you think about it, it's obvious that raid-1 will read faster than it |
9 |
> writes - it has to write two copies while it only reads one. |
10 |
|
11 |
Yes. The same is true for RAID10, since it has to also write two |
12 |
copies of everything. |
13 |
|
14 |
> |
15 |
> Likewise, raids 5 and 6 will be slower writing than reading - for a |
16 |
> normal read it only reads the data disks, but when writing it has to |
17 |
> write (and calculate!) parity as well. |
18 |
|
19 |
Yes, but with any of the striped modes (0, 5, 6, 10) there is an |
20 |
additional issue. Writes have to generally be made in entire stripes, |
21 |
so if you overwrite data in-place in units smaller than an entire |
22 |
stripe, then the entire stripe needs to first be read, and then it can |
23 |
be overwritten again. This is an absolute requirement if there is |
24 |
parity involved. If there is no parity (RAID 0,10) then an |
25 |
implementation might be able to overwrite part of a stripe in place |
26 |
without harming the rest. |
27 |
|
28 |
> |
29 |
> A raid 1 should read data faster than a lone disk. A raid 5 or 6 should |
30 |
> read noticeably faster because it's reading across more than one disk. |
31 |
|
32 |
More-or-less. RAID 1 is going to generally benefit from lower latency |
33 |
because reads can be divided across mirrored copies (and there could |
34 |
be more than one replica). Any of the striped modes are going to be |
35 |
the same as a single disk on latency, but will have much greater |
36 |
bandwidth. That bandwidth gain applies to both reading and writing, |
37 |
as long as the data is sequential. |
38 |
|
39 |
This is why it is important to understand your application. There is |
40 |
no one "best" RAID implementation. They all have pros and cons |
41 |
depending on whether you care more about latency vs bandwidth and also |
42 |
read vs write. |
43 |
|
44 |
And of course RAID isn't the only solution out there for this stuff. |
45 |
Distributed filesystems also have pros and cons, and often those have |
46 |
multiple modes of operation on top of this (usually somewhat mirroring |
47 |
the options available for RAID but across multiple hosts). |
48 |
|
49 |
For general storage I'm using zfs with raid1 pairs of disks (the pool |
50 |
can have multiple pairs), and for my NAS for larger-scale media/etc |
51 |
storage I'm using lizardfs. I'd use ceph instead in any kind of |
52 |
enterprise setup, but that is much more RAM-hungry and I'm cheap. |
53 |
|
54 |
-- |
55 |
Rich |