Hi!
I have a Bro cluster on a 2xE5-2660 with 64GB RAM. The traffic comes via
a 10Gbps port of an Intel 82599ES (it's not cleared from FIN_storms and
so on). I use Netmap with the Netmap ixgbe drivers and lb. lb and Bro
workers pinned to cores. The interface offloads are turned off (ethtool
-L eth0 combined 1; ... ). The basic installation (1xlb, 20 cores for
Bro) begins drops packets hard near 7Gbps (~900Kpps). It looks like the
bottleneck was the lb (100% the core usage and it shows drops). Lite lb
optimisation ( I unify the loops in sym_hash_fn) doesn't help too much.
I try the following lb pyramid:
lb -i eth0 -B10000 -p bro:2 -o 60
lb -i netmap:bro}0 -B10000 -p broa:13 -o 60
lb -i netmap:bro}1 -B10000 -p brob:13 -o 60
and :
...
[worker-1]
type=worker
host=localhost
interface=netmap::broa
lb_method=custom
lb_procs=13
pin_cpus=7,8,9,10,11,12,13,14,15,16,17,18,19
[worker-2]
type=worker
host=localhost
interface=netmap::brob
lb_method=custom
lb_procs=13
pin_cpus=22,23,24,25,26,27,28,29,30,31,32,33,34
in node.cfg .
Now I can get up to 8Gbps (~1000Kpps) without a lot of losses.
But is it possible to get full 10Gbps on a port with netmap and lb on a
host like mine? How to offload lb?
Thanks,
Alexander Zatserkovnyy