1 |
On 30/09/2013 12:32, Joost Roeleveld wrote: |
2 |
> On Monday 30 September 2013 10:01:32 Alan McKinnon wrote: |
3 |
>> On 30/09/2013 06:14, Walter Dnes wrote: |
4 |
>>> If the udev people had made "net ifnames=0" the default, and allowed |
5 |
>>> |
6 |
>>> the small percentage of multi-nic machine admins to set "net.ifnames=1", |
7 |
>>> this would not have been an issue. Some corner case exotic setups |
8 |
>>> require complex solutions... no ifs/ands/ors/buts. All the complaining |
9 |
>>> you hear is from the other 99% who's setup worked just fine with the |
10 |
>>> simple solution, suddenly finding the complex solution rammed down their |
11 |
>>> throats. |
12 |
>> |
13 |
>> No, that is just plain wrong. |
14 |
>> |
15 |
>> Having interfaces on a multi-nic host come up as ethX where X is a |
16 |
>> mostly random number is just so broken it beggars belief. Trust me, it |
17 |
>> is zero fun when it happens and what makes it even worse if you have no |
18 |
>> warning at all beforehand. |
19 |
> |
20 |
> I trust you, but on my multi-nic systems, I found a better solution :) |
21 |
> As I use Xen to virtualize my systems and as I don't want to have multiple |
22 |
> network cables running side-by-side, I started using VLANs. |
23 |
> |
24 |
> I know have all the NICs names eth1,eth2,...ethn. |
25 |
> I throw them all as a bonded network device: bond0 (the other ends go into a |
26 |
> switch supporting bonding network ports) |
27 |
> then on top of that, I have VLANs with distinctive names (lan, dmz, guest, |
28 |
> vm,...) and link these as required to different Xen-domains. |
29 |
> |
30 |
> When the network names get renamed suddenly to the "non-predictive" scheme, my |
31 |
> system refuses to boot. |
32 |
> Before that, I would use mac-addresses to link ethx devices to names that make |
33 |
> sense to me. (see above for the names) |
34 |
|
35 |
|
36 |
The worst case that comes to mind was a three zone netflow collector |
37 |
plus the first nic on our management range. |
38 |
|
39 |
If you're familiar with old netflow versions you'll know it is UDP from |
40 |
the router and is touchy about addresses. So we had incoming netflow |
41 |
from three ranges each hitting a dedicated nic and this all worked |
42 |
marvellously for years and years. |
43 |
|
44 |
One day after a routine maintenance window the box came up with all 4 |
45 |
nics scrambled and who knows what was now assigned to what. Forget ssh |
46 |
to log in and fix it - nothing was listening. That took very senior |
47 |
sysadmins on site to deal with, the regular maintenance guy was in way |
48 |
over his head. |
49 |
|
50 |
Business were OK with losing 15 minutes billing and stats data in a |
51 |
maintenance window. They were definitely not OK with losing several |
52 |
hours of it because someone thought assigning names on a |
53 |
non-deterministic discovery order was a good idea. |
54 |
|
55 |
One thing about Dell hardware - you always know exactly what each nic is |
56 |
connected to on the motherboard so with that info the new names are |
57 |
predictable (consistent is actually the better term). Using MAC |
58 |
addresses for the same purposes is clunky and unwieldy, the MACs have to |
59 |
be recorded somewhere and you still don't know which MAC goes with which |
60 |
physical socket. With bus numbers you do know. |
61 |
|
62 |
|
63 |
-- |
64 |
Alan McKinnon |
65 |
alan.mckinnon@×××××.com |