On Mon, 28 Nov 2016 13:21:41 +0100 Jesper Dangaard Brouer <bro...@redhat.com>
wrote:
> On Mon, 28 Nov 2016 11:52:38 +0100 Paolo Abeni <pab...@redhat.com> wrote:
> >
> > > > [2] like [1], but using the minimum number of flows to saturate the
> > > > user space
> > > > sink, that is 1 flow for the old kernel and 3 for the patched one.
> > > > the tput increases since the contention on the rx lock is low.
> > > > [3] like [1] but using a single flow with both old and new kernel. All
> > > > the
> > > > packets land on the same rx queue and there is a single ksoftirqd
> > > > instance
> > > > running
[...]
> >
> > We also used connected socket for test[3], with relative little
> > difference (the tput increased for both unpatched and patched kernel,
> > and the difference was roughly the same).
>
> When I use connected sockets (RX side) and ip_early_demux enabled, I do
> see a performance boost for recvmmsg. With these patches applied,
> forced ksoftirqd on CPU0 and udp_sink on CPU2, pktgen single flow
> sending size 1472 bytes.
>
> $ sysctl net/ipv4/ip_early_demux
> net.ipv4.ip_early_demux = 1
>
> $ grep -H . /proc/sys/net/core/{r,w}mem_max
> /proc/sys/net/core/rmem_max:1048576
> /proc/sys/net/core/wmem_max:1048576
>
> # taskset -c 2 ./udp_sink --count $((10**7)) --port 9 --repeat 1
> # ns pps cycles
> recvMmsg/32 run: 0 10000000 462.51 2162095.23 1853
> recvmsg run: 0 10000000 536.47 1864041.75 2150
> read run: 0 10000000 492.01 2032460.71 1972
> recvfrom run: 0 10000000 553.94 1805262.84 2220
>
> # taskset -c 2 ./udp_sink --count $((10**7)) --port 9 --repeat 1 --connect
> # ns pps cycles
> recvMmsg/32 run: 0 10000000 405.15 2468225.03 1623
> recvmsg run: 0 10000000 548.23 1824049.58 2197
> read run: 0 10000000 489.76 2041825.27 1962
> recvfrom run: 0 10000000 466.18 2145091.77 1868
>
> My theory is that by enabling connect'ed RX socket, the ksoftirqd gets
> faster (no fib_lookup) and is no-longer a bottleneck. This is
> confirmed by nstat.
Paolo asked me to do a test with small packets with pktgen, and I was
actually surprised by the result.
# taskset -c 2 ./udp_sink --count $((10**7)) --port 9 --repeat 1 --connect
recvMmsg/32 run: 0 10000000 426.61 2344076.59 1709 17098657328
recvmsg run: 0 10000000 533.49 1874449.82 2138 21382574965
read run: 0 10000000 470.22 2126651.13 1884 18846797802
recvfrom run: 0 10000000 513.74 1946499.83 2059 20591095477
Notice how recvMmsg/32, got slower with 124kpps (2468225 pps -> 2344076 pps).
I was expecting it to get faster, given we just established udp_sink
was the bottleneck, and smaller packet should mean less copy of bytes
to userspace (copy_user_enhanced_fast_string). (With nstat I observe
ksoftirq is again the bottleneck).
Looking at perf diff of CPU2 (baseline=64Bytes) we do see an increase
of copy_user_enhanced_fast_string. More interestingly we see a
decrease in the locking cost when using big packets (see ** below)
# Event 'cycles:ppp'
#
# Baseline Delta Shared Object Symbol
# ........ ....... ................ .........................................
#
15.09% +0.33% [kernel.vmlinux] [k] copy_msghdr_from_user
12.36% +21.89% [kernel.vmlinux] [k] copy_user_enhanced_fast_string
8.65% -0.63% [kernel.vmlinux] [k] udp_process_skb
7.33% -1.88% [kernel.vmlinux] [k] __skb_try_recv_datagram_batch
** 7.12% -6.66% [kernel.vmlinux] [k] udp_rmem_release **
** 6.71% -6.52% [kernel.vmlinux] [k] _raw_spin_lock_bh **
6.35% +1.36% [kernel.vmlinux] [k] __free_page_frag
4.39% +0.29% [kernel.vmlinux] [k] copy_msghdr_to_user_gen
2.87% -1.52% [kernel.vmlinux] [k] skb_release_data
2.60% +0.14% [kernel.vmlinux] [k] __put_user_4
2.27% -2.18% [kernel.vmlinux] [k] __sk_mem_reduce_allocated
2.11% +0.08% [kernel.vmlinux] [k] cmpxchg_double_slab.isra.68
1.90% +2.40% [kernel.vmlinux] [k] __slab_free
1.73% +0.20% [kernel.vmlinux] [k] __udp_recvmmsg
1.62% -1.62% [kernel.vmlinux] [k] intel_idle
1.52% +0.22% [kernel.vmlinux] [k] copy_to_iter
1.20% -0.03% [kernel.vmlinux] [k] import_iovec
1.14% +0.05% [kernel.vmlinux] [k] rw_copy_check_uvector
0.80% -0.04% [kernel.vmlinux] [k] recvmmsg_ctx_to_user
0.75% -0.69% [kernel.vmlinux] [k] __local_bh_enable_ip
0.71% +0.18% [kernel.vmlinux] [k] skb_copy_datagram_iter
0.70% -0.07% [kernel.vmlinux] [k] recvmmsg_ctx_from_user
0.67% +0.08% [kernel.vmlinux] [k] kmem_cache_free
0.56% +0.42% [kernel.vmlinux] [k] udp_process_msg
0.48% +0.05% [kernel.vmlinux] [k] skb_release_head_state
0.46% [kernel.vmlinux] [k] lapic_next_deadline
0.36% [kernel.vmlinux] [k] __switch_to
0.34% -0.03% [kernel.vmlinux] [k] consume_skb
0.32% -0.05% [kernel.vmlinux] [k] skb_consume_udp
The perf diff from CPU0, also show less lock congestion:
# Event 'cycles:ppp'
#
# Baseline Delta Shared Object Symbol
# ........ ....... ................ .........................................
#
11.04% -3.02% [kernel.vmlinux] [k] __udp_enqueue_schedule_skb
9.98% +2.16% [mlx5_core] [k] mlx5e_handle_rx_cqe
7.23% -1.85% [kernel.vmlinux] [k] udp_v4_early_demux
3.90% +0.73% [kernel.vmlinux] [k] build_skb
3.85% -1.77% [kernel.vmlinux] [k] udp_queue_rcv_skb
3.83% +0.02% [kernel.vmlinux] [k] sock_def_readable
** 3.26% -3.19% [kernel.vmlinux] [k] queued_spin_lock_slowpath **
2.99% +0.55% [kernel.vmlinux] [k] __build_skb
2.97% +0.11% [kernel.vmlinux] [k] __udp4_lib_rcv
** 2.87% -1.39% [kernel.vmlinux] [k] _raw_spin_lock **
2.67% +0.60% [kernel.vmlinux] [k] ip_rcv
2.65% +0.61% [kernel.vmlinux] [k] __netif_receive_skb_core
2.64% +0.79% [ip_tables] [k] ipt_do_table
2.37% +0.37% [kernel.vmlinux] [k] read_tsc
2.26% +0.52% [mlx5_core] [k] mlx5e_get_cqe
2.11% -1.15% [kernel.vmlinux] [k] __sk_mem_raise_allocated
2.10% +0.37% [kernel.vmlinux] [k] __rcu_read_unlock
2.04% +0.67% [mlx5_core] [k] mlx5e_alloc_rx_wqe
1.86% +0.40% [kernel.vmlinux] [k] inet_gro_receive
1.57% +0.11% [kernel.vmlinux] [k] kmem_cache_alloc
1.53% +0.28% [kernel.vmlinux] [k] _raw_read_lock
1.53% +0.25% [kernel.vmlinux] [k] dev_gro_receive
1.38% -0.18% [kernel.vmlinux] [k] udp_gro_receive
1.19% +0.37% [kernel.vmlinux] [k] __rcu_read_lock
1.14% +0.31% [kernel.vmlinux] [k] _raw_read_unlock
1.14% +0.12% [kernel.vmlinux] [k] ip_rcv_finish
1.13% +0.20% [kernel.vmlinux] [k] __udp4_lib_lookup
1.05% +0.16% [kernel.vmlinux] [k] ktime_get_with_offset
0.94% +0.38% [kernel.vmlinux] [k] ip_local_deliver_finish
0.91% +0.22% [kernel.vmlinux] [k] do_csum
0.86% -0.04% [kernel.vmlinux] [k] ipv4_pktinfo_prepare
0.84% +0.05% [kernel.vmlinux] [k] sk_filter_trim_cap
0.84% +0.20% [kernel.vmlinux] [k] ip_local_deliver
0.84% +0.19% [kernel.vmlinux] [k] udp4_gro_receive
--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
Author of http://www.iptv-analyzer.org
LinkedIn: http://www.linkedin.com/in/brouer
