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® 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® 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® 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® 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® 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® Ethernet, visit http://communities.intel.com/community/wired
