Re: [E1000-devel] e1000e heavy packet drop

[Ambrus W. <amb...@gm...>]

Hi,

update on this issue.

I have installed an earlier release of Ubuntu (8.04, long term support
version) and the earliest driver (0.2.9.5), but unfortunately this
didn't help either. The drops are still there, although it seems that
without any kernel tuning packet drop is below 1% until 1Gbps (on 4
interfaces). This is not much, but better than nothing...

If you have any ideas, please share with me...

Thanks a lot,
Ambrus

2009/7/21 Ambrus Wagner <amb...@gm...>:
> Hi,
>
> we are developing a packet capture and filtering application for Linux
> based on a quad port Intel Pro/1000 PT adapter.
>
> A similar setup is already working in production environment:
> - kernel version:  2.6.18.2
> - driver: e1000e 0.2.9.5
>
> In the above case we are not experiencing any packet loss, the input
> is approximately 1.2Gbps spread across the four ports. There is no
> transmission.
>
> We are building another setup with the following config:
> - same adapter (Intel Pro/1000 PT quad port, have tried with dual port
> as well, same result)
> - kernel version: 2.6.30.1
> - driver: e1000e 1.0.2.5
>
> The input traffic is about the same, with disastrous results
> (extracted from ifconfig output after a couple of minutes):
>
> eth2 RX packets:157863020 errors:0 dropped:129568261 overruns:0 frame:0
> eth3 RX packets:158295124 errors:0 dropped:129074182 overruns:0 frame:0
> eth9 RX packets:158134206 errors:0 dropped:128962842 overruns:0 frame:0
> eth10 RX packets:158184566 errors:0 dropped:128846931 overruns:0 frame:0
>
> The ethtool -S output shows that the majority of the drop comes from
> rx_missed_errors:
>
>     rx_packets: 167976786
>     tx_packets: 4
>     rx_bytes: 54452791812
>     tx_bytes: 1384
>     rx_broadcast: 167976071
>     tx_broadcast: 4
>     rx_multicast: 715
>     tx_multicast: 0
>     rx_errors: 0
>     tx_errors: 0
>     tx_dropped: 0
>     multicast: 715
>     collisions: 0
>     rx_length_errors: 0
>     rx_over_errors: 0
>     rx_crc_errors: 0
>     rx_frame_errors: 0
>     rx_no_buffer_count: 23842322
>     rx_missed_errors: 138693966
>
> Meanwhile ksoftirq uses 100% CPU. The system by the way is a Core2 Duo.
>
> The configuration of the module and the interfaces is as follows:
> - copybreak=0
> - promiscuous mode
> - mtu 1500
> - RX ring set to 4096
>
> We have already tried without copybreak, without mtu setting, lots of
> RX ring sizes, lots of interrupt throttling settings (runs at 3 now,
> but have tried everything from 1 and 2 to 100, 1000, 10000, etc.)
>
> In the previous setup CPU is over 95% idle, with ksoftirq consuming
> almost no cycles.
>
> We have tried kernel 2.6.30 already (based on a post somewhere), but
> we have not tried to go back to 2.6.18 on the same system as we don't
> really want to reinstall it from scratch if not necessary. Also, the
> driver version 0.2.9.5 does not compile any more on this kernel, so we
> are pretty much stuck here, although I believe this setup should work
> anyway.
>
> Any help is very much appreciated, we are totally clueless by now.
>
> Thanks,
> Ambrus
>
> --
> Wagner, Ambrus
> e-mail: amb...@gm...
>



-- 
Wagner, Ambrus
e-mail: amb...@gm...