1 |
I am going to answer multiple points below: |
2 |
|
3 |
|
4 |
On 1/3/20 11:36 pm, Daniel Frey wrote: |
5 |
> On 3/1/20 6:33 AM, Rich Freeman wrote: |
6 |
>> On Sun, Mar 1, 2020 at 2:13 AM William Kenworthy <billk@×××××××××.au> |
7 |
>> wrote: |
8 |
>>> |
9 |
>>> Keep in mind that rpi are not the only cheap, capable arm hardware out |
10 |
>>> there. |
11 |
>>> |
12 |
>> |
13 |
I am in Oz, delivery from HardKernel (the South Korean company behind |
14 |
the Odroid line) takes ~1 week. Shipping is mostly via FedEx, who are a |
15 |
bit pricy and for me means a 30m drive to get it from a delivery centre |
16 |
as I cant be home to take delivery - though the last delivery used a |
17 |
different, much more flexible delivery company. |
18 |
|
19 |
lizardfs and moosefs are very similar (originally a fork) - I went with |
20 |
moosefs as the community was better a few months ago - but lizardfs |
21 |
sounds like its getting back on track after some infighting (moosefs |
22 |
apparently went through the same thing) I find the moosefs |
23 |
documentation and community help are adequate, but ultimately they are |
24 |
are a commercial project with a free taster offering hence some |
25 |
limitations in design for the community version (such as single |
26 |
master/no shadow masters) that lizardfs doesn't have. I will move to |
27 |
lizardfs when I am satisfied they have their act together because of |
28 |
this - having a single master means taking the whole cluster offline |
29 |
when the master needs maintenance or fails which is painful |
30 |
|
31 |
For those wanting to run a lot of drives on a single host - that defeats |
32 |
the main advantage of using a chunkserver based filesystem - |
33 |
redundancy. Its far more common to have a host fail than a disk drive. |
34 |
Losing the major part of your storage in one go means the cluster is |
35 |
effectively dead - hence having a lot of completely separate systems is |
36 |
much more reliable - yes, I did try having 4 sata drives on an atom |
37 |
board and found it was easy to justify two more HC2's for the |
38 |
reliability. (note, its not the effect of the atom boards reliability I |
39 |
am pointing out, but the effect on the whole storage system of losing |
40 |
such a large percentage of capacity in one go - think maintenance, fail |
41 |
to startup etc. - its more common than you may think) |
42 |
|
43 |
Dale: there is no single reference - I just designed on the fly and did |
44 |
what was necessary to move from my existing KVM/Qemu architecture on a |
45 |
two powerfull intel systems, each with 8TB storage to an lxc container |
46 |
based system backed by moosefs. It uses an odroid arm based n2 for lxc, |
47 |
an arm based xu4 for management and software compilation (its not that |
48 |
powerfull, but is adequate), the HC2's (very similar to the xu4 - same |
49 |
arm architecture) and an H2 to run the mfsmaster and I am moving to |
50 |
ansible for management. Overall, I am finding less maintenance due to |
51 |
better backups and better reliability, considerably lower power |
52 |
consumption while overall performance seems better. |
53 |
|
54 |
Gotchas: |
55 |
|
56 |
1. I originally got the 4gb ram N2 to run the mfsmaster software - tests |
57 |
were excellent - until I added multiple copies of a mailserver with |
58 |
nearly 20 years of history - hit swap and slowed to a crawl. So I now |
59 |
have an intel based Odroid H2 with 32GB ram - currently is using ~4.2GB |
60 |
ram for 7TB of 20TB used, but has hit over 32GbB and well into swap |
61 |
while converging. One admin in our local LUG is using 4xHC2 with the |
62 |
master running on one of the HC2's as a media server with no problems - |
63 |
its millions of small files added ina short time that causes the grief - |
64 |
there is a formula in the moosefs documentation allowing resources to be |
65 |
estimated - on the mailing list I saw a mention of a data centre using |
66 |
something like 150G ram and having problems! |
67 |
|
68 |
2. The Odroid HC2 has a single sata port and a single usb port - I have |
69 |
5 of them, two have a sata + a usb3 drive attached. I did try (on the |
70 |
N2) using multiple usb3 drives on the internal hub - disaster with way |
71 |
too much traffic through the hub - don't do it! |
72 |
|
73 |
3. Storage options are an SD card (the faster the better - swap on an SD |
74 |
card ... sucks!) or eMMC (5xfaster than even a good SD card.) for the N2 |
75 |
and xu4 - HC2's can only do sdcard, or sata. The H2 can do sd card, |
76 |
sata, eMMC, or m2 NVME - this last really flies! Note that almost all |
77 |
arm system max out at 2 to 4GB of ram, so swap is usually needed for |
78 |
safety - depending on OOM killer resource management on a SAN type |
79 |
storage system is asking for corruption like I saw in one recent gentoo |
80 |
email. |
81 |
|
82 |
4. xu4/HC2 are 32 bit arm v7, the N2 is 64 bit and runs aarch64 nicely - |
83 |
I copied my rpi images across repurposed them (hooray for |
84 |
emerge/portage!). I do not use, or have done any work on their |
85 |
graphics/multimedia capabilities |
86 |
|
87 |
5. I want to move to all gentoo-sources kernels - the xu4/HC2's are |
88 |
still on the odroid kernel until I get around to it. The N2 was a lot |
89 |
of work, but ultimately successful, the H2 is standard amd64 EFI. |
90 |
|
91 |
|
92 |
Have fun! |
93 |
|
94 |
BillK |