currently bge(4) allocates 3.5M (~0.7%) of kernel memory per interface
that is never used unless jumbo frames are configured.
you can do the math to see that this is a horrible waste.
and that is in addition to normal mbuf clusters allocated for
normal frames (that is another 512k or so).
here is the diff that only allocates those huge buffers for jumbos
once the mtu has been raised high enough.
please test esp if you are using jumbo frames.
paranoic mickey (my employers have changed but, the name has remained)