Dual NIC clients

[Quazz]

I also tried adding option tftp-server-name X.X.X.X; to my dhcpd config file, but no luck.

The correct option is next-server X.X.X.X

That’s probably already correctly set judging by your post.

Your problem may be resolved by trunk, maybe not, but you’d need to install newer Ubuntu version for it.

[tag]

@Quazz - Thanks for the reply.

You’d be correct in assuming that next-server is set as per default by the FOG installation script.

If i understand you correctly, your suggestion of trunking would enable the client to connect to the TFTP server on either link? I’d rather not, as the other link is only 100Mbps on some cheap Cisco 2960 SI series.

Ideally the clients would choose the correct link matching the network ID of the next-server.

[Quazz]

@tag You can define wich NIC to use in FOG, so no, it would only use that NIC. Everything else will depend a lot on your network setup so I can’t comment on that.

[tag]

@Quazz - Thanks again.

How do you mean define which NIC to use in FOG? I see only options for setting the server side NIC… Am I missing something here?

[Quazz]

@tag Sorry, it’s still morning, things got mixed up in my head.

I think this can only be solved by directing traffic through network control, but I don’t really know enough about it to comment.

[tag]

@Quazz said in Dual NIC clients:

Sorry, it’s still morning, things got mixed up in my head.

LOL, tank up on your morning stimulant of choice…

[tag]

OK - did some thinking. It times out because there is no default gateway set on the secondary link. Setting that it will connect. The problem is, how do I know which network it chooses? I’m getting inconsistent transfer speeds now, average of 5GB/min versus 225MB/min - apparently depending on which NIC it connects through.

I should mention that inter-VLAN routing is enabled on the layer 3 switch of the primary network. Removing the secondary network from the static route list or pulling the physical link kills it again - this time at trying to send an inventory before deploying.

If I pull the power on the secondary network they will all deploy at high speeds.
With the secondary network on (and inter-VLAN routing), some will deploy normally, others slowly - apparently arbitrarily, as the same machines will act differently from task to task.
It would seem the kernel arbitrarily sets which NIC is eth0 from boot to boot? That would perhaps explain why it would appear to use different NICs.

If I pull the plug on the secondary network while deploying it stops deploying until plugged back in. So it’s using the “wrong” NIC…

Anyone ever see something like this?

[Quazz]

@tag Is it arbitrarily? If you try the same PC over and over does it report different speeds?

I think it has to do with the order your system reports the NICs in. If it reports the slower NIC first it will use that. Perhaps that’s something you can alter on your end?

[tag]

@Quazz
Yes, sometimes an image will deploy at 5GB/min - a few minutes later when trying again with the same client it will only deploy at a fraction of that speed…

Otherwise I agree; I too believe it has to do with the order of the NICs.

I tried swapping the cables at first to see if it just chose one specific hardware ID first, but that was not the case.

Since then I’ve done quite a few test deployments on the same eigth machines. Mostly they’ll deploy slowly, but every now and again one will run on the faster NIC, which can be verified by pulling the cables and seeing which one makes it pause.

I’m not saying it’s arbitrary, but I don’t see a pattern so it seems arbitrary to me.

[Quazz]

@tag Are you using quick image?

Might be possible to tell it to use the faster NIC by registering them, assigning the faster NIC as primary NIC and deploying in that manner, but I really can’t be certain on that, I was kind of hoping someone else would chime on in on this, heh.

[tag]

@Quazz
No, they are registered and deployed through tasks. Actually it’s not a question of one NIC being faster than the other - they’re more or less identical mobo dual NICs - as it is the link speed. They’re 100Mbps switch ports on a 100Mbps trunk to the layer 3 switch. It was never intended for large data transfers - just remote access and so on.

The primary MAC in the host registration is the faster link, so that has no effect, I’m afraid. These are 1Gbps switch ports for the clients and two ports in ether channel for the server.

[Wayne Workman]

@tag said in Dual NIC clients:

If i understand you correctly, your suggestion of trunking would enable the client to connect to the TFTP server on either link?

No, the developmental version of fog is called “fog trunk”.

[Wayne Workman]

@Quazz do we know of any linux kernel arguments that specify only using a particular nic, or disabling a particular nic?

Maybe even a custom kernel could be an answer? Or custom init?

[Quazz]

@Wayne-Workman The problem is, what arguments can we use to differentiate? It sounds like they’re basically identical NICs connecting to different network outlets, probably getting inconsistent names as well. (based on his results)

https://www.kernel.org/doc/Documentation/kernel-parameters.txt

Scroll down to grcan.enable0

Looks like those options are useful?

[Wayne Workman]

@Quazz very nice. I read it’s description, but the one below it caught my eye:

grcan.select= [HW] Select which physical interface to use. Format: 0 | 1 Default: 0

So perhaps try this for the host’s kernel arguments (web gui -> host management -> desired host -> kernel arguments)

grcan.select=1

See what happens?

[tag]

@Quazz and @Wayne-Workman

Thanks for the suggestions.

I tried playing around with grcan.enable0=[0|1] and grcan.enable1=[0|1] as well as grcan.select=[0|1] but none had any effect. The kernel continues to choose the slower link in most cases. Seemed promising, though…

What I do notice for the first time, though, is that whatever NIC is not chosen is disabled. I hadn’t noticed as I can’t see the backs of the boxes very well. Here it is also obvious that the active NIC changes on occasion, as the LEDs die on the disabled NIC.

[Wayne Workman]

@tag I think you’re going to have to build a custom init. You can change the fog.upload and fog.download scripts. The idea would be to use shell script to determine which interface is on the right network, and then disable the other interface (or enable it). It should be pretty simple.

How experienced are you with shell scripting?

Also, here’s a link on how to unpack and re-pack the inits:
https://wiki.fogproject.org/wiki/index.php?title=Build_FOG_file_system_with_BuildRoot

I’m willing to help do this - but I wouldn’t have time until tonight to mess with it.

[tag]

@Wayne-Workman

Thanks for the reply.

Seems kind of inflexible, though… The same init is used for all, right? We even have some clients with three NICs at other locations… If it has to take various hw scenarios into account, it might take some fancy scripting.

I know some basic scripting but nothing really fancy.

The only way to determine the correct interface would be to filter on IP, as I see it.

So maybe a list of interfaces and then for each ethX in the list:

#!/bin/csh

set nwid = X.X.X
set list = (eth0 eth1)

foreach eth ($list)
  set ip = `ifconfig $eth | grep inet | awk '{print $2}' | sed 's/addr://' | cut -c-10 `
  if ($ip == $nwid) then
      ifup $eth else
      ifdown $eth
  endif
end

That would in my case get the network ID of the correct network and other disimilar outputs from the other interfaces in the list which could then be compared to a set value of the correct network ID. Based on that comparison you could then turn on or off the interfaces.

I’m sure someone else could do something a lot niftier.

I haven’t tested any of this and it might screw up if the number of interfaces actually present is different from the number in the list.

[Wayne Workman]

@tag You’ve got the right idea - but that specific code is inflexible.

Tonight I’ll put something together that will take the IPs, and the subnet mask, and calculate the subnet ID and use that for comparison.

[tag]

@Wayne-Workman
Thanks again.

Yes, that code will only work on the specific network defined in $nwid and if the kernel names the interfaces ethX and probably only if the number of interfaces match that particular piece of hw…

Mighty nice of you to help me out here… Appreciate it.

Thanks.