PF_RING server load for comparison. But I sure netmap better then PF_RING many 
times and I do something wrong :)

I can provide perf top output too:
  8.23%  [ixgbe]           [k] ixgbe_netmap_rxsync
  7.97%  [ixgbe]           [k] ixgbe_poll
  7.88%  [kernel]          [k] irq_entries_start
  4.15%  [kernel]          [k] do_gettimeofday
  4.12%  [netmap_lin]      [k] netmap_poll
  4.00%  [kernel]          [k] add_interrupt_randomness
  3.03%  [netmap_lin]      [k] nm_kr_put
  2.32%  [netmap_lin]      [k] nm_kr_tryget
  2.27%  [kernel]          [k] native_read_tsc
  1.88%  [kernel]          [k] _raw_spin_lock_irqsave
  1.25%  [kernel]          [k] arch_local_irq_restore
  1.11%  [kernel]          [k] __schedule
  1.09%  [kernel]          [k] __netif_receive_skb
  1.07%  [kernel]          [k] timekeeping_get_ns
  1.05%  [kernel]          [k] arch_local_irq_save
  1.04%  [kernel]          [k] _raw_spin_unlock_irqrestore
  0.99%  [netmap_lin]      [k] nm_rxsync_prologue
  0.98%  [kernel]          [k] do_raw_spin_lock
  0.87%  [kernel]          [k] __alloc_skb
  0.69%  [kernel]          [k] arch_local_irq_enable
  0.69%  [kernel]          [k] arch_local_irq_restore
  0.63%  [kernel]          [k] __wake_up
  0.62%  [kernel]          [k] idle_cpu
  0.62%  [kernel]          [k] getnstimeofday
  0.59%  [kernel]          [k] select_estimate_accuracy
  0.59%  [kernel]          [k] __cache_free.isra.41
  0.57%  [kernel]          [k] handle_edge_irq
  0.56%  [ixgbe]           [k] test_and_set_bit
  0.56%  [kernel]          [k] do_sys_poll
  0.52%  [kernel]          [k] __do_softirq
  0.52%  [kernel]          [k] handle_irq_event_percpu
  0.51%  [kernel]          [k] do_IRQ
  0.51%  [kernel]          [k] atomic_inc
  0.51%  [kernel]          [k] rcu_exit_nohz
  0.51%  [kernel]          [k] net_rx_action
  0.50%  [kernel]          [k] timerqueue_add
  0.50%  [kernel]          [k] dev_gro_receive
  0.50%  [kernel]          [k] tick_nohz_stop_sched_tick

After some NIC tuning kernels load become smaller but still big for 200kpps: 

Sorry for wrong information.

In 8 thread mode:
../examples//pkt-gen -i eth3 -f rx -p 8

All my cpus loaded like this: 
https://netmap.googlecode.com/issues/attachment?aid=300000000&name=netmap_overlo
ad.png&token=ABZ6GAdSCcuQyFqynZp9JRBluLBPzySfUw%3A1414578954266&inline=1

Hi Pavel,
we tried to investigate this problem. Maybe the cause is the Interrupt Throttle 
Rate (ITR).

Can you try to set a big value for the rx-side?
sudo ethtool -C eth3 rx-usecs 1000

Cheers,
Stefano

Hello, folks!

I tried with FreeBSD 10 and hit same issue.

My configuration:
VirtualBox, VM with 4 CPUs i7 3635QM, 2.4Ghz
Intel PRO1000/MT Server, bridge mode

I wrote this code for tests: 
https://github.com/FastVPSEestiOu/fastnetmon/blob/master/tests/netmap.cpp

Then I start flooding my server with hping3: sudo hping3 --flood --udp 
192.168.0.2

hping3 generated about 20 000 packets per server but this load complately 
overload my server with netmap.

Average server load on system was about 30%. It's really really slow even for 
pcap :(

Any news? :(

Looks like this issue related with my hand made broken patch of Sourceforge's 
Intel ixgbe driver.