Hi Donald,

> Once again to clarify your concern is:
> 
> TX system -> RX system (TCP) 9.5 Gb/s
> TX system -> RX system (UDP / FC off) 5.6 Gb/s


Sorry... not exactly, but I guess I found the reason why we can't 
understand each other. When I say: "on TX side I get 9.5 Gb/s" I mean 
that I measure the network traffic through the interface on the TX 
system  (I use dstat for that).

I try to clarify further my numbers:

All tests are from TX system to RX system

UDP (no FC):
through eth0 on TX system: 9.5 Gb/s
through eth0 on RX system: 5.6 Gb/s

UDP (FC):
through eth0 on TX system 5.6 Gb/s
through eth0 on RX system 5.6 Gb/s

TCP (no FC):
through eth0 on TX system 5.6 Gb/s
through eth0 on RX system 5.6 Gb/s

TCP (FC):
through eth0 on TX system 5.6 Gb/s
through eth0 on RX system 5.6 Gb/s

FC does not improve my throughput (but of course eliminates 
rx_missed_errors).

> As for the meaning of the rx_missed_errors I can at least help you with that. 
>  I imagined you looked at the code but it's the summation of an array of 
> registers called MPC(), one for each receive FIFO.  It's incremented when 
> Packets are missed due to insufficient space to store that packet.  This 
> might be due to bandwidth issue with the bus IO (which is why I was asking 
> about the slot you have the NIC's in) or too few buffers allocated.  If I 
> remember correctly you already tried that.  Or in our case a transmitter 
> blasting out packets as faster than we can receive them

Ok. As I wrote before this is my network buffer configuration:

net.core.wmem_max = 16777216
net.core.wmem_default = 8388608
net.core.rmem_max = 16777216
net.core.rmem_default = 8388608
net.ipv4.tcp_rmem = 4096 262144 16777216
net.ipv4.tcp_wmem = 4096 262144 16777216
net.ipv4.tcp_mem = 4096 8388608 16777216
net.core.optmem_max = 524288
net.core.netdev_max_backlog = 200000

AFAIK, these are the only parameters I can set.

Thanks for your patience :-)

Mirko

>> -----Original Message-----
>> From: Mirko Corosu [mailto:[email protected]]
>> Sent: Wednesday, February 24, 2010 2:56 AM
>> To: [email protected]
>> Subject: Re: [E1000-devel] Intel 2598EB 10-Gigabit AT dropped rx packet
>>
>> Hi Donald,
>>
>>
>>> One thing I did notice was that the Receive side has 8 rx queues while
>> the Transmit has 16 rx queues.  You mentioned these were identical machines
>> could Hyper-Threading be off on the Receiving machine?
>> Yes, it was my fault. I forgot to disable hyperthreading on TX server.
>> Now it is disabled, but the problem still remains.
>>
>>> Also I'm still a little confused about how your test is set up.  If you
>> have four netperf's running on, as you called it, your Tx machine and your
>> Rx is only running netserver then I would only expect heavy traffic from Tx
>> -> Rx.  Or am I missing some thing here?
>> I try to summarize my tests.
>>
>> For now I am testing only mono-directional TCP and UDP stream (from TX
>> server to RX server directly ). What I cannot understand is the low
>> throughput that I'm getting. The TCP tests shows a throughput of only
>> 5.6 Gb/s. During UDP tests, with flow control disabled, I measured on TX
>> side a throughput of about 9.5 Gb/s (so the card can send at wire speed)
>> but on RX side I see only 5.6 Gb/s. I attribute this problem to the high
>> number of rx_missed_errors but I cannot figure out what is the real
>> cause (I don't know even what rx_missed_errors actually mean....).
>>
>> Thanks
>>
>> Mirko
>>
>>>> -----Original Message-----
>>>> From: Mirko Corosu [mailto:[email protected]]
>>>> Sent: Tuesday, February 23, 2010 3:30 AM
>>>> To: Skidmore, Donald C
>>>> Cc: [email protected]
>>>> Subject: Re: [E1000-devel] Intel 2598EB 10-Gigabit AT dropped rx packet
>>>>
>>>> Hi Donald,
>>>>
>>>> Ok, sorry for the lack of informations: I'm not used to interact with
>>>> developers. :-)
>>>>
>>>>
>>>>> How were you using netperf in your tests?  What tests (I assume
>>>> [UDP|TCP]_STREAM) and how many instances of netperf were running each
>> side?
>>>> I'm running four instances of UDP_STREAM or TCP_STREAM tests on the tx
>>>> side, with different packet sizes (from 1000 to 9000 byte) and a single
>>>> netserver on the rx side.
>>>>
>>>>> From the spec I got off the web on the R710 I can see they have either
>> (2
>>>> PCIe x8 and 2 PCIe x4) or (1 PCIe x16 and 2 PCIe x4 ).  I was just
>>>> concerned you might be plugged into one of the x4 slots.
>>>>
>>>> I have  2 PCIe x8 on the so called "Riser 2" and 2 PCIe x4 on the "Riser
>>>> 1".   On each server the Intel cards are plugged into the Riser 2 slots.
>>>>
>>>>>>>  - Did you look at the dmesg for anything relative?
>>>>>> I see no error or warning messages on both side.
>>>>> It's not just error or warning messages I was interested in, although
>>>> they would be interesting too. :)  We also printk a message that
>> displays
>>>> the type of PCIe link we have.  This would answer the question I posed
>>>> above about what slot the cards were in.
>>>> On rx side:
>>>>
>>>> [r...@grids2 ~]# dmesg |grep ixgbe
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 8, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:bf
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> None
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 8, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:bf
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> None
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 8, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:bf
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 8, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:bf
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> None
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 8, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:bf
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 8, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:bf
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>>
>>>>
>>>>
>>>> On tx side:
>>>>
>>>> [r...@client20 ~]# dmesg |grep ixgbe
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 16, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: No DCA provider found.  Please start ioatdma
>>>> for DCA functionality.
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:e0
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> None
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Down
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 16, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: No DCA provider found.  Please start ioatdma
>>>> for DCA functionality.
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:e0
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 16, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: No DCA provider found.  Please start ioatdma
>>>> for DCA functionality.
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:e0
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>> ixgbe: eth0: ixgbe_remove: complete
>>>> ixgbe: Intel(R) 10 Gigabit PCI Express Network Driver - version
>>>> 2.0.44.14-NAPI
>>>> ixgbe: 0000:07:00.0: ixgbe_init_interrupt_scheme: Multiqueue Enabled: Rx
>>>> Queue count = 16, Tx Queue count = 1
>>>> ixgbe: eth0: ixgbe_probe: No DCA provider found.  Please start ioatdma
>>>> for DCA functionality.
>>>> ixgbe: eth0: ixgbe_probe: (PCI Express:2.5Gb/s:Width x8)
>> 00:1b:21:4b:c8:e0
>>>> ixgbe: eth0: ixgbe_probe: MAC: 1, PHY: 2, PBA No: d79893-017
>>>> ixgbe: eth0: ixgbe_probe: Internal LRO is enabled
>>>> ixgbe: eth0: ixgbe_probe: Intel(R) 10 Gigabit Network Connection
>>>> ixgbe: eth0: ixgbe_change_mtu: changing MTU from 1500 to 9000
>>>> ixgbe: eth0: ixgbe_watchdog_task: NIC Link is Up 10 Gbps, Flow Control:
>>>> RX/TX
>>>>
>>>>
>>>>
>>>>> All that said we would expect to see rx_missed_errors with flow control
>>>> disabled.  Since nothing would be stopping packets from overrunning the
>>>> buffer.  Your test seems to be agreeing with this as the number of
>>>> rx_missed_errors goes down with TCP (like you said) most likely due to
>> TCP
>>>> sliding window.
>>>>> You don't see this issue with flow control enabled, right?
>>>> Right, but the throughput remains about 5.6 Gb/s and a can't figure out
>>>> where the bottleneck could be....
>>>>
>>>> Thanks a lot
>>>>
>>>> Mirko
>>>>
>>>>>> -----Original Message-----
>>>>>> From: Mirko Corosu [mailto:[email protected]]
>>>>>> Sent: Saturday, February 20, 2010 10:50 AM
>>>>>> To: Skidmore, Donald C
>>>>>> Cc: [email protected]
>>>>>> Subject: Re: [E1000-devel] Intel 2598EB 10-Gigabit AT dropped rx
>> packet
>>>>>> Hi Donald,
>>>>>>
>>>>>> Thank you for your reply.
>>>>>>
>>>>>>> I have a couple additional questions
>>>>>>>  - What were you running to do this test? (i.e. netperf, pktgen?)
>>>>>> netperf-2.4.2-1.el5.rf
>>>>>>
>>>>>>>  - What throughput were you seeing?
>>>>>> On tx side i see almost 10Gb/s, on rx side I see about 5.6 Gb/s
>>>>>>
>>>>>>>  - Are both ports plugged into x8 PCIe slots
>>>>>> Yes, both servers are Dell R710 with the same hardware configuration
>>>>>>
>>>>>>>  - Did you look at the dmesg for anything relative?
>>>>>> I see no error or warning messages on both side.
>>>>>>
>>>>>>>  - Do you see this issue only with UDP?
>>>>>> If I send a TCP stream I see on both side a throughput of about 5.54
>>>>>> Gb/s and less rx_missed_errors on rx side:
>>>>>>
>>>>>> NIC statistics:
>>>>>>      rx_packets: 1768101
>>>>>>      tx_packets: 235384
>>>>>>      rx_bytes: 15937257320
>>>>>>      tx_bytes: 15704254
>>>>>>      lsc_int: 2
>>>>>>      tx_busy: 0
>>>>>>      non_eop_descs: 7072222
>>>>>>      rx_errors: 0
>>>>>>      tx_errors: 0
>>>>>>      rx_dropped: 0
>>>>>>      tx_dropped: 0
>>>>>>      multicast: 1
>>>>>>      broadcast: 0
>>>>>>      rx_no_buffer_count: 0
>>>>>>      collisions: 0
>>>>>>      rx_over_errors: 0
>>>>>>      rx_crc_errors: 0
>>>>>>      rx_frame_errors: 0
>>>>>>      rx_fifo_errors: 0
>>>>>>      rx_missed_errors: 1539
>>>>>>      tx_aborted_errors: 0
>>>>>>      tx_carrier_errors: 0
>>>>>>      tx_fifo_errors: 0
>>>>>>      tx_heartbeat_errors: 0
>>>>>>      tx_timeout_count: 0
>>>>>>      tx_restart_queue: 0
>>>>>>      rx_long_length_errors: 0
>>>>>>      rx_short_length_errors: 0
>>>>>>      tx_tcp4_seg_ctxt: 0
>>>>>>      tx_tcp6_seg_ctxt: 0
>>>>>>      tx_flow_control_xon: 0
>>>>>>      rx_flow_control_xon: 0
>>>>>>      tx_flow_control_xoff: 0
>>>>>>      rx_flow_control_xoff: 0
>>>>>>      rx_csum_offload_good: 1768098
>>>>>>      rx_csum_offload_errors: 0
>>>>>>      tx_csum_offload_ctxt: 11
>>>>>>      low_latency_interrupt: 0
>>>>>>      alloc_rx_page_failed: 0
>>>>>>      alloc_rx_buff_failed: 0
>>>>>>      lro_aggregated: 1760417
>>>>>>      lro_flushed: 375277
>>>>>>      lro_recycled: 1012791
>>>>>>      rx_no_dma_resources: 0
>>>>>>      hw_rsc_count: 0
>>>>>>      rx_flm: 0
>>>>>>      fdir_match: 0
>>>>>>      fdir_miss: 0
>>>>>>      tx_queue_0_packets: 235384
>>>>>>      tx_queue_0_bytes: 15704254
>>>>>>      rx_queue_0_packets: 2
>>>>>>      rx_queue_0_bytes: 120
>>>>>>      rx_queue_1_packets: 8
>>>>>>      rx_queue_1_bytes: 786
>>>>>>      rx_queue_2_packets: 301752
>>>>>>      rx_queue_2_bytes: 2719959436
>>>>>>      rx_queue_3_packets: 562919
>>>>>>      rx_queue_3_bytes: 5074118826
>>>>>>      rx_queue_4_packets: 7
>>>>>>      rx_queue_4_bytes: 726
>>>>>>      rx_queue_5_packets: 15
>>>>>>      rx_queue_5_bytes: 1526
>>>>>>      rx_queue_6_packets: 313901
>>>>>>      rx_queue_6_bytes: 2829476778
>>>>>>      rx_queue_7_packets: 589497
>>>>>>      rx_queue_7_bytes: 5313699122
>>>>>>
>>>>>> I guess it's because of TCP flow control.
>>>>>>
>>>>>> I remember that the same behavior was triggered by enabling flow
>> control
>>>>>> on both cards (ethtool -A eth0 rx on tx on autoconf on) but I have to
>>>>>> wait till Monday to perform another run to test that configuration.
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Mirko
>>>>>>
>>>>>>> What I'm wondering is if this NIC was receiving faster than it could
>>>>>> process them.
>>>>>>> Thanks,
>>>>>>> -Don Skidmore <[email protected]>
>>>>>>>
>>>>>>>
>>>>>>>> -----Original Message-----
>>>>>>>> From: Mirko Corosu [mailto:[email protected]]
>>>>>>>> Sent: Friday, February 19, 2010 5:30 AM
>>>>>>>> To: [email protected]
>>>>>>>> Subject: [E1000-devel] Intel 2598EB 10-Gigabit AT dropped rx packet
>>>>>>>>
>>>>>>>> Dear all,
>>>>>>>>
>>>>>>>> I'm new to this list, if this isn't the right place to post this
>>>> issue,
>>>>>>>> please let me know.
>>>>>>>>
>>>>>>>> I am testing two Intel 82598EB 10-Gigabit AT network card mounted on
>>>> two
>>>>>>>> Dell R710 server with Red Hat Enterprise Linux 5.4 installed. The
>> two
>>>>>>>> NICs are directly connected each other (no switch in between).   I
>> am
>>>>>>>> experiencing a problem on the receive side of the connection: I'm
>>>>>> sending
>>>>>>>> an UDP stream after having disabled flow control on each card, and
>>>> about
>>>>>>>> the 60% of the transmitted packets are dropped with a large number
>> of
>>>>>>>> rx_missed_errors. This is what ifconfig and ethtool -S show:
>>>>>>>>
>>>>>>>> [r...@grids2]# ethtool -S eth0
>>>>>>>> NIC statistics:
>>>>>>>>      rx_packets: 1761794
>>>>>>>>      tx_packets: 32
>>>>>>>>      rx_bytes: 14748667944
>>>>>>>>      tx_bytes: 5746
>>>>>>>>      lsc_int: 2
>>>>>>>>      tx_busy: 0
>>>>>>>>      non_eop_descs: 6510828
>>>>>>>>      rx_errors: 0
>>>>>>>>      tx_errors: 0
>>>>>>>>      rx_dropped: 0
>>>>>>>>      tx_dropped: 0
>>>>>>>>      multicast: 2
>>>>>>>>      broadcast: 2
>>>>>>>>      rx_no_buffer_count: 0
>>>>>>>>      collisions: 0
>>>>>>>>      rx_over_errors: 0
>>>>>>>>      rx_crc_errors: 0
>>>>>>>>      rx_frame_errors: 0
>>>>>>>>      rx_fifo_errors: 0
>>>>>>>>      rx_missed_errors: 1148382
>>>>>>>>      tx_aborted_errors: 0
>>>>>>>>      tx_carrier_errors: 0
>>>>>>>>      tx_fifo_errors: 0
>>>>>>>>      tx_heartbeat_errors: 0
>>>>>>>>      tx_timeout_count: 0
>>>>>>>>      tx_restart_queue: 0
>>>>>>>>      rx_long_length_errors: 0
>>>>>>>>      rx_short_length_errors: 0
>>>>>>>>      tx_tcp4_seg_ctxt: 0
>>>>>>>>      tx_tcp6_seg_ctxt: 0
>>>>>>>>      tx_flow_control_xon: 0
>>>>>>>>      rx_flow_control_xon: 0
>>>>>>>>      tx_flow_control_xoff: 0
>>>>>>>>      rx_flow_control_xoff: 0
>>>>>>>>      rx_csum_offload_good: 28
>>>>>>>>      rx_csum_offload_errors: 0
>>>>>>>>      tx_csum_offload_ctxt: 11
>>>>>>>>      low_latency_interrupt: 0
>>>>>>>>      alloc_rx_page_failed: 0
>>>>>>>>      alloc_rx_buff_failed: 0
>>>>>>>>      lro_aggregated: 8
>>>>>>>>      lro_flushed: 8
>>>>>>>>      lro_recycled: 0
>>>>>>>>      rx_no_dma_resources: 0
>>>>>>>>      hw_rsc_count: 0
>>>>>>>>      rx_flm: 0
>>>>>>>>      fdir_match: 0
>>>>>>>>      fdir_miss: 0
>>>>>>>>      tx_queue_0_packets: 32
>>>>>>>>      tx_queue_0_bytes: 5746
>>>>>>>>      rx_queue_0_packets: 2
>>>>>>>>      rx_queue_0_bytes: 120
>>>>>>>>      rx_queue_1_packets: 2
>>>>>>>>      rx_queue_1_bytes: 120
>>>>>>>>      rx_queue_2_packets: 1761762
>>>>>>>>      rx_queue_2_bytes: 14748664800
>>>>>>>>      rx_queue_3_packets: 0
>>>>>>>>      rx_queue_3_bytes: 0
>>>>>>>>      rx_queue_4_packets: 0
>>>>>>>>      rx_queue_4_bytes: 0
>>>>>>>>      rx_queue_5_packets: 14
>>>>>>>>      rx_queue_5_bytes: 1452
>>>>>>>>      rx_queue_6_packets: 7
>>>>>>>>      rx_queue_6_bytes: 726
>>>>>>>>      rx_queue_7_packets: 7
>>>>>>>>      rx_queue_7_bytes: 726
>>>>>>>>
>>>>>>>> [r...@grids2]# ifconfig eth0
>>>>>>>>
>>>>>>>> eth0      Link encap:Ethernet  HWaddr 00:1B:21:4B:C8:BF
>>>>>>>>           inet addr:10.100.100.2  Bcast:10.100.100.255
>>>>>> Mask:255.255.255.0
>>>>>>>>           inet6 addr: fe80::21b:21ff:fe4b:c8bf/64 Scope:Link
>>>>>>>>           UP BROADCAST RUNNING MULTICAST  MTU:9000  Metric:1
>>>>>>>>           RX packets:1761795 errors:0 dropped:1148382 overruns:0
>>>> frame:0
>>>>>>>>           TX packets:39 errors:0 dropped:0 overruns:0 carrier:0
>>>>>>>>           collisions:0 txqueuelen:1000
>>>>>>>>           RX bytes:14748668004 (13.7 GiB)  TX bytes:9332 (9.1 KiB)
>>>>>>>>
>>>>>>>> Each interface has the MTU set to 9000, the driver is the one
>> provided
>>>>>>>> by Dell support (ixgbe 2.0.44.14-NAPI) with Rx multiqueue enabled
>> and
>>>>>>>> InterruptThrottleRate set to 8000 (but I tried 16000 and 32000 too).
>>>>>>>> Also I tried to set FdirPballoc parameter to 0, 1 and 2 with no
>>>> success.
>>>>>>>> Single CPU load is no more than 10% with 8000 interrupt/s:
>>>>>>>>
>>>>>>>> [r...@grids2 ~]# mpstat -P ALL 2
>>>>>>>>
>>>>>>>> 05:16:43 PM  CPU   %user   %nice    %sys %iowait    %irq   %soft
>>>> %steal
>>>>>>>> %idle    intr/s
>>>>>>>> 05:16:45 PM  all    0.00    0.00    0.06    0.00    0.06    1.19
>>>> 0.00
>>>>>>>>    98.69   9018.00
>>>>>>>> 05:16:45 PM    0    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00   1002.50
>>>>>>>> 05:16:45 PM    1    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00      1.50
>>>>>>>> 05:16:45 PM    2    0.00    0.00    0.50    0.00    0.50    9.45
>>>> 0.00
>>>>>>>>    89.55   8007.00
>>>>>>>> 05:16:45 PM    3    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00      1.00
>>>>>>>> 05:16:45 PM    4    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00      1.00
>>>>>>>> 05:16:45 PM    5    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00      3.00
>>>>>>>> 05:16:45 PM    6    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00      1.00
>>>>>>>> 05:16:45 PM    7    0.00    0.00    0.00    0.00    0.00    0.00
>>>> 0.00
>>>>>>>>   100.00      1.00
>>>>>>>>
>>>>>>>> I tried to configure smp affinity to assign the IRQ of each
>>>>>>>> rx-queue to a different CPU, with no effect on the amount of dropped
>>>>>>>> packets. I also verified that if I connect the NIC on a 10GE switch
>>>> and
>>>>>>>> send more than one stream from different servers, the card uses
>>>>>>>> different rx-queues but the problem still exists.
>>>>>>>>
>>>>>>>> I modified the /etc/sysctl.conf file parameters adding:
>>>>>>>>
>>>>>>>> net.core.wmem_max = 16777216
>>>>>>>> net.core.wmem_default = 8388608
>>>>>>>> net.core.rmem_max = 16777216
>>>>>>>> net.core.rmem_default = 8388608
>>>>>>>> net.ipv4.tcp_rmem = 4096 262144 16777216
>>>>>>>> net.ipv4.tcp_wmem = 4096 262144 16777216
>>>>>>>> net.ipv4.tcp_mem = 4096 8388608 16777216
>>>>>>>> net.core.optmem_max = 524288
>>>>>>>> net.core.netdev_max_backlog = 200000
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Some other details:
>>>>>>>>
>>>>>>>> - CPU: dual quad core Intel L5520  @ 2.27GHz (Hyper-Threading
>>>> disabled)
>>>>>>>> - RAM: 24 GB (12 X 2GB)
>>>>>>>> - Slot PCI: PCIe gen.2 x8
>>>>>>>> - Kernel linux: 2.6.18-128.1.1.el5 and 2.6.18-164.2.1.el5
>>>>>>>>
>>>>>>>> Any suggestions?
>>>>>>>> Thank you very much in advance.
>>>>>>>>
>>>>>>>> Mirko Corosu
>>>>>>>>
>>>>>>>> --
>>>>>>>>
>>>>>>>> --------------------------------------------------------------------
>> --
>>>> --
>>>>>> --
>>>>>>>> Mirko Corosu
>>>>>>>> Network and system administrator
>>>>>>>> Computing Center
>>>>>>>> Istituto Nazionale Fisica Nucleare
>>>>>>>> Via Dodecaneso 33
>>>>>>>> 16146 Genova, Italy
>>>>>>>> http://www.ge.infn.it
>>>>>>>> --------------------------------------------------------------------
>> --
>>>> --
>>>>>> --
>>>>>>>> --------------------------------------------------------------------
>> --
>>>> --
>>>>>> ---
>>>>>>>> ---
>>>>>>>> Download Intel&#174; Parallel Studio Eval
>>>>>>>> Try the new software tools for yourself. Speed compiling, find bugs
>>>>>>>> proactively, and fine-tune applications for parallel performance.
>>>>>>>> See why Intel Parallel Studio got high marks during beta.
>>>>>>>> http://p.sf.net/sfu/intel-sw-dev
>>>>>>>> _______________________________________________
>>>>>>>> E1000-devel mailing list
>>>>>>>> [email protected]
>>>>>>>> https://lists.sourceforge.net/lists/listinfo/e1000-devel
>>>>>>>> To learn more about Intel&#174; Ethernet, visit
>>>>>>>> http://communities.intel.com/community/wired
>>>>>> --
>>>>>>
>>>>>> ----------------------------------------------------------------------
>> --
>>>> --
>>>>>> Mirko Corosu
>>>>>> Istituto Nazionale Fisica Nucleare
>>>>>> Via Dodecaneso 33
>>>>>> 16146 Genova
>>>>>> www.ge.infn.it
>>>>>> Phone +39 010 3536361
>>>>>> ----------------------------------------------------------------------
>> --
>>>> --
>>>>>> ----------------------------------------------------------------------
>> --
>>>> --
>>>>>> Se tutto sembra venirti incontro, probabilmente sei nella corsia
>>>> sbagliata.
>>>>>> ----------------------------------------------------------------------
>> --
>>>> --
>>>>
>>>> --
>>>>
>>>> ------------------------------------------------------------------------
>> --
>>>> Mirko Corosu
>>>> Network and system administrator
>>>> Computing Center
>>>> Istituto Nazionale Fisica Nucleare
>>>> Via Dodecaneso 33
>>>> 16146 Genova, Italy
>>>> http://www.ge.infn.it
>>>> ------------------------------------------------------------------------
>> --
>>>>
>>>>
>>>> ------------------------------------------------------------------------
>> --
>>>> Se tutto sembra venirti incontro, probabilmente sei nella corsia
>> sbagliata.
>>>> ------------------------------------------------------------------------
>> --
>>
>> --
>> --------------------------------------------------------------------------
>> Mirko Corosu
>> Network and system administrator
>> Computing Center
>> Istituto Nazionale Fisica Nucleare
>> Via Dodecaneso 33
>> 16146 Genova, Italy
>> http://www.ge.infn.it
>> --------------------------------------------------------------------------
>>
>> --------------------------------------------------------------------------
>> Se tutto sembra venirti incontro, probabilmente sei nella corsia sbagliata.
>> --------------------------------------------------------------------------
>>
>>
>>
>> ---------------------------------------------------------------------------
>> ---
>> Download Intel&#174; Parallel Studio Eval
>> Try the new software tools for yourself. Speed compiling, find bugs
>> proactively, and fine-tune applications for parallel performance.
>> See why Intel Parallel Studio got high marks during beta.
>> http://p.sf.net/sfu/intel-sw-dev
>> _______________________________________________
>> E1000-devel mailing list
>> [email protected]
>> https://lists.sourceforge.net/lists/listinfo/e1000-devel
>> To learn more about Intel&#174; Ethernet, visit
>> http://communities.intel.com/community/wired


-- 
--------------------------------------------------------------------------
Mirko Corosu
Network and system administrator
Computing Center
Istituto Nazionale Fisica Nucleare
Via Dodecaneso 33
16146 Genova, Italy
http://www.ge.infn.it
--------------------------------------------------------------------------

--------------------------------------------------------------------------
Se tutto sembra venirti incontro, probabilmente sei nella corsia sbagliata.
--------------------------------------------------------------------------


------------------------------------------------------------------------------
Download Intel&#174; Parallel Studio Eval
Try the new software tools for yourself. Speed compiling, find bugs
proactively, and fine-tune applications for parallel performance.
See why Intel Parallel Studio got high marks during beta.
http://p.sf.net/sfu/intel-sw-dev
_______________________________________________
E1000-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/e1000-devel
To learn more about Intel&#174; Ethernet, visit 
http://communities.intel.com/community/wired

Reply via email to