1 |
On Sat, Jul 31, 2021 at 11:05 PM William Kenworthy <billk@×××××××××.au> wrote: |
2 |
> |
3 |
> On 31/7/21 9:30 pm, Rich Freeman wrote: |
4 |
> > |
5 |
> > I'd love server-grade ARM hardware but it is just so expensive unless |
6 |
> > there is some source out there I'm not aware of. It is crazy that you |
7 |
> > can't get more than 4-8GiB of RAM on an affordable arm system. |
8 |
> Checkout the odroid range. Same or only slightly $$$ more for a much |
9 |
> better unit than a pi (except for the availability of 8G ram on the pi4) |
10 |
|
11 |
Oh, they have been on my short list. |
12 |
|
13 |
I was opining about the lack of cheap hardware with >8GB of RAM, and I |
14 |
don't believe ODROID offers anything like that. I'd be happy if they |
15 |
just took DDR4 on top of whatever onboard RAM they had. |
16 |
|
17 |
My SBCs for the lizardfs cluster are either Pi4s or RockPro64s. The |
18 |
Pi4 addresses basically all the issues in the original Pis as far as |
19 |
I'm aware, and is comparable to most of the ODroid stuff I believe (at |
20 |
least for the stuff I need), and they're still really cheap. The |
21 |
RockPro64 was a bit more expensive but also performs nicely - I bought |
22 |
that to try playing around with LSI HBAs to get many SATA drives on |
23 |
one SBC. |
24 |
|
25 |
I'm mainly storing media so capacity matters more than speed. At the |
26 |
time most existing SBCs either didn't have SATA or had something like |
27 |
1-2 ports, and that means you're ending up with a lot of hosts. Sure, |
28 |
it would perform better, but it costs more. Granted, at the start I |
29 |
didn't want more than 1-2 drives per host anyway until I got up to |
30 |
maybe 5 or so hosts just because that is where you see the cluster |
31 |
perform well and have decent safety margins, but at this point if I |
32 |
add capacity it will be to existing hosts. |
33 |
|
34 |
> Tried ceph - run away fast :) |
35 |
|
36 |
Yeah, it is complex, and most of the tools for managing it created |
37 |
concerns that if something went wrong they could really mess the whole |
38 |
thing up fast. The thing that pushed me away from it was reports that |
39 |
it doesn't perform well only a few OSDs and I wanted something I could |
40 |
pilot without buying a lot of hardware. Another issue is that at |
41 |
least at the time I was looking into it they wanted OSDs to have 1GB |
42 |
of RAM per 1TB of storage. That is a LOT of RAM. Aside from the fact |
43 |
that RAM is expensive, it basically eliminates the ability to use |
44 |
low-power cheap SBCs for all the OSDs, which is what I'm doing with |
45 |
lizardfs. I don't care about the SBCs being on 24x7 when they pull a |
46 |
few watts each peak, and almost nothing when idle. If I want to |
47 |
attach even 4x14TB hard drives to an SBC though it would need 64GB of |
48 |
RAM per the standards of Ceph at the time. Good luck finding a cheap |
49 |
low-power ARM board that has 64GB of RAM - anything that even had DIMM |
50 |
slots was something crazy like $1k at the time and at that point I |
51 |
might as well build full PCs. |
52 |
|
53 |
It seems like they've backed off on the memory requirements, maybe, |
54 |
but I'd want to check on that. I've seen stories of bad things |
55 |
happening when the OSDs don't have much RAM and you run into a |
56 |
scenario like: |
57 |
1. Lose disk, cluster starts to rebuild. |
58 |
2. Lose another disk, cluster queues another rebuild. |
59 |
3. Oh, first disk comes back, cluster queues another rebuild to |
60 |
restore the first disk. |
61 |
4. Replace the second failed disk, cluster queues another rebuild. |
62 |
|
63 |
Apparently at least in the old days all the OSDs had to keep track of |
64 |
all of that and they'd run out of RAM and basically melt down, unless |
65 |
you went around adding more RAM to every OSD. |
66 |
|
67 |
With LizardFS the OSDs basically do nothing at all but pipe stuff to |
68 |
disk. If you want to use full-disk encryption then there is a CPU hit |
69 |
for that, but that is all outside of Lizardfs and dm-crypt at least is |
70 |
reasonable. (zfs on the other hand does not hardware accelerate it on |
71 |
SBCs as far as I can tell and that hurts.) |
72 |
|
73 |
> I improved performance and master memory |
74 |
> requirements considerably by pushing the larger data sets (e.g., Gib of |
75 |
> mail files) into a container file stored on MFS and loop mounted onto |
76 |
> the mailserver lxc instance. Convoluted but very happy with the |
77 |
> improvement its made. |
78 |
|
79 |
Yeah, I've noticed as you described in the other email memory depends |
80 |
on number of files, and it depends on having it all in RAM at once. |
81 |
I'm using it for media storage mostly so the file count is modest. I |
82 |
do use snapshots but only a few at a time so it can handle that. |
83 |
While the master is running on amd64 with plenty of RAM I do have |
84 |
shadow masters set up on SBCs and I do want to be able to switch over |
85 |
to one if something goes wrong, so I want RAM use to be acceptable. |
86 |
It really doesn't matter how much space the files take up - just now |
87 |
many inodes you have. |
88 |
|
89 |
-- |
90 |
Rich |