Hello,

I have several systems in a cluster ( > 65 nodes and counting) using either bonded ethernet (802.3ad), or mellanox infiniband. I need to delay the boot process until the interfaces are actually configured and running. It can take up to 20 seconds though from the time the kernel module is loaded to the interface being able to function. I have troubleshot this but it pretty much boils down to "the connection takes time to come up".

In the case of the bonded ethernet, I am connected to a cisco switch, and the network admin assures me this time is normal. In the case of the infiniband, it looks like the same thing: the infiniband fabric just takes time to come up.

The result is services don''t start because the network connections are not present, or they timeout and fail.

Other than inserting a sleep statement in the run control scripts, is there a method of handling this? I would like the system to pause only long enough for the interfaces to be operable.

I have this "trouble" with centos5, centos 6, ubuntu 10.04 lts, and ubuntu 12.04lts.

--jason

A conditional wait statement in a while loop is going to be your best bet.

(I've never had bonded ethernet interfaces take anywhere near that long to come up. What bonding mode are you using?)

Hello,

The bonding mode is 4 with miimod set to 400. The connections are all to the same cisco switch.

I probably should clarify at least one thing though: dmesg shows that the links themselves are coming up just fine, it's the time it takes for tcp/ip to be functional. While the address is assigned to bond0, it takes a while for the switch to actually begin routing it. The time it takes reminds me of cisco's "spanning-tree portfast" setting being off.

I haven't been looking at it as slow because I have the same trouble with infiniband connections coming up.

--jason