So daqd's problem was apparently the bad/non-running c1sup model. The c1sup model, which I reported on attempting to get running in 7097, was not running because there were no available CPUs on the c1sus FE machine. This was due to my stupid undercounting of the number of CPUs. Anyway, for reasons I don't understand, this was causing daqd to segfault. Removing c1sup from c1sus "fixed" the problem.
Alex agreed that daqd should definitely not be segfaulting in this circumstance. It's still unclear exactly what daqd was looking at that was causing it to crash.
I'm going to move c1sup to c1iscex, which has a lot of spare CPUs.
I spoke too soon. It's still segfaulting, but at a different place. Alex and I are looking into it.
But another mystery solved is the cause of all the network slowness: the daqd core dump. When daqd segfaults it dumps it's core, which can typically be >4G, to /opt/rtcds/caltech/c1/target/fb/core. This is of course an NFS mount from linux1, so it's dumping 4G on the network, which not surprisingly clogs the network.