Similarly to what has been done earlier with other actions [1][2], this
series tries to improve the performance of 'csum' tc action, removing a
spinlock in the data plane. Patch 1 lets act_csum use per-CPU counters;
patch 2 removes spin_{,un}lock_bh() calls from act() method.
test procedure:
# ip link add name eth1 type dummy
# ip link set dev eth1 up
# tc qdisc add dev eth1 root handle 1: prio
# tc filter add dev eth1 parent 1: matchall action csum udp
# for n in 2 4 6 8 10 12 14 16; do
> ./pktgen_bench_xmit_mode_queue_xmit.sh -v -s 64 -t $n -n 1000000 -i eth1
> done
test results:
$n | avg. pps/core | avg. pps/core | delta
| (without patch) | (with patch) | (%)
---+-----------------+---------------+------
2 | 484915 | 547716 | 13
4 | 209551 | 254439 | 21
6 | 143901 | 164695 | 14
8 | 112423 | 127821 | 14
10 | 91134 | 102950 | 13
12 | 75374 | 85499 | 13
14 | 64586 | 73426 | 14
16 | 56635 | 64111 | 13
references:
[1] http://www.spinics.net/lists/netdev/msg334760.html
[2] https://www.spinics.net/lists/netdev/msg465862.html
Davide Caratti (2):
net/sched: act_csum: use per-core statistics
net/sched: act_csum: don't use spinlock in the fast path
net/sched/act_csum.c | 17 ++++++-----------
1 file changed, 6 insertions(+), 11 deletions(-)
--
2.13.6
