Created attachment 170931 [details] server tcpdump showing the bug Hi all, We found an unexpected behavior in the applications which we are using that appears to be a bug in the TCP algorithm. Following tcpdumps we detected that a second unnecessary retransmission timeout (RTO) occurs after a first valid RTO. This occurs only in a very particular situation: when there are 2 packets pending to be sent at the moment the first retransmission occurs (and also in the context of our application which operates in a very low latency network, and which its only traffic is receiving a single request message and sending a ~12KB response). This situation occurs very often in our context and because the application operates at very low latency a RTO impacts severely the performance. More details of the application communication and tcpdumps with the bug explanation is copied below. Is this a proper bug or is there something we are missing expected TCP behavoir? Also, is there is this a known way to avoid this unexpected behaviour? Thank you very much, Regards, Matias Environment OS: SLC 6 $ uname -a Linux <pcName> 2.6.32-504.3.3.el6.x86_64 #1 SMP Wed Dec 17 09:22:39 CET 2014 x86_64 x86_64 x86_64 GNU/Linux TCP configuration: full configuration from /proc/sys/net/ipv4/* pasted at the bottom. tcp_base_mss=512 tcp_congestion_control=cubic tcp_sack=1 Application communication description: We are using 2 applications which communicate using TCP: a client and around 200 server application. The client sends a request message of 188B (including headers) to all servers and waits for a response of all of them. The client does not send any other message until the response of all servers is received. Each server upon receiving the request, it sends a 12KB response (which is obviously splitted into several TCP packets). Because there are 200 servers responding at almost the same moment (with a total of ~2.4MB) some buffers in the network may overflow generating drops and retransmissions. When there are no drops (thanks to control application that limits the requests sent) the latency to receive all messages from all servers is ~20ms. If there is a drop of one or more TCP fragments then the latency goes to near ~200ms (this is because of the minimum RTO of 200ms hardcoded in the kernel). Even if this is 10 times higher it is more or less under acceptable for the application. The bug creates a second consecutive retransmission so the latency when this occurs goes to 600ms (200ms of the first RTO + 400ms of the second unexpected RTO), which is out of the limits that the application can handle (60 times higher). Bug detailed description: The unexpected behavior appears in the server applications when TCP needs to retransmit drops packets. It appears in all server applications at a quite a high frequency. The bug appears only when the server detected a drop (by a RTO after 200ms) and at that moment it is still pending to receive the ACK for 2 packets. In that case, after 200ms of sending all packets, the RTO triggers the retransmission of the first packet, then the ACK for that packet is received but the second packet is not retransmitted at that moment. After another 400ms another RTO is triggered and that second packet is retransmitted and ACKed. To our understanding this second retransmission should not occur. The expected behaviour is that the second packet is retransmitted right after receiving the ACK for the first retransmitted packet. Also this unexpected second RTO occurs only if there are 2 pending packets at the moment of the first RTO. If there is one packet to retransmit for more than 2, the behaviour is as expected, all packets are retransmitted and ACKed after the first RTO (there is no second RTO). Below the explanation and a section of a tcpdump recorded in one of the server applications showing the unexpected behaviour. Frame #170: request is received (at T 0) Frames #171-#173: response is sent splitted into several TCP packets. From seq=204273 to seq=216289. Frame #171 and #172 are recorded by tcpdump as a single packet but is probably several real packets as the MSS is 1460 bytes and it shows a lenght higher than that (this is probably caused because NICs support segmentation offloads, which means, that the NIC joins the segments together and pushes it to the host’s TCP stack a single segment. This is why tcpdump sees it as a segment of higher length). Frames #173-#177: ACKs for some of the sent packets is received. Last seq acknowledged is seq=442797 (there is still 1796 bytes to be sent, which is 2 TCP packets). Frame #178: At T 207ms a packet is retransmitted. This is the first retransmission, which makes total sense as the ACKs for 2 packets were not received after 200ms. Because of the RTO the TCP internal state should be updated to duplicate the RTO (so it should be 400ms now). Also the CWND should be reduced to 1. Frame #179: ACK for the retransmitted packet is received. The internal state of TCP should be update to duplicate the CWND because of slow start (so should be set now to 2). RTO is not updated because calculation of RTO is based only in packets which were not retransmitted. At this point we would expect that the pending packet should be retransmitted, but this does not occur. After receiving an ACK the CWDN should allow more packets to be sent, but there is no data sent by the server (and consequently it receives nothing). Frame #180: at T 613ms (aprox ~400ms after the last received ACK) the last packet is retransmitted. This is what creates a 600ms latency which is 60 times the expected and 6 times higher if the bug would not be present. Frame #181: ACK for the last packet is received. Frame #182: a new request is received.. No. Time Source Destination Protocol RTO Length Info 170 *REF* DCM ROS TCP 118 47997 > 41418 [PSH, ACK] Seq=1089 Ack=204273 Win=10757 Len=64 171 0.000073 ROS DCM TCP 5894 41418 > 47997 [ACK] Seq=204273 Ack=1153 Win=58 Len=5840 172 0.000080 ROS DCM TCP 5894 41418 > 47997 [ACK] Seq=210113 Ack=1153 Win=58 Len=5840 173 0.000083 ROS DCM TCP 390 41418 > 47997 [PSH, ACK] Seq=215953 Ack=1153 Win=58 Len=336[Packet size limited during capture] 174 0.003901 DCM ROS TCP 60 47997 > 41418 [ACK] Seq=1153 Ack=207193 Win=10757 Len=0 175 0.004270 DCM ROS TCP 60 47997 > 41418 [ACK] Seq=1153 Ack=211573 Win=10768 Len=0 176 0.004649 DCM ROS TCP 60 47997 > 41418 [ACK] Seq=1153 Ack=213033 Win=10768 Len=0 177 0.004835 DCM ROS TCP 66 [TCP Dup ACK 176#1] 47997 > 41418 [ACK] Seq=1153 Ack=213033 Win=10768 Len=0 SLE=214493 SRE=215953 178 0.207472 ROS DCM TCP 0.207389000 1514 [TCP Retransmission] 41418 > 47997 [ACK] Seq=213033 Ack=1153 Win=58 Len=1460 179 0.207609 DCM ROS TCP 60 47997 > 41418 [ACK] Seq=1153 Ack=215953 Win=10768 Len=0 180 0.613472 ROS DCM TCP 0.613389000 390 [TCP Retransmission] 41418 > 47997 [PSH, ACK] Seq=215953 Ack=1153 Win=58 Len=336[Packet size limited during capture] 181 0.613622 DCM ROS TCP 60 47997 > 41418 [ACK] Seq=1153 Ack=216289 Win=10768 Len=0 182 0.615189 DCM ROS TCP 118 47997 > 41418 [PSH, ACK] Seq=1153 Ack=216289 Win=10768 Len=64 Full TCP configuration: for f in /proc/sys/net/ipv4/* ;do confName=$(basename "$f") ; echo -n "$confName=" >> /logs/tpu_TCP_config.txt ; cat "$f" >> /logs/tpu_TCP_config.txt ;done cipso_cache_bucket_size=10 cipso_cache_enable=1 cipso_rbm_optfmt=0 cipso_rbm_strictvalid=1 icmp_echo_ignore_all=0 icmp_echo_ignore_broadcasts=1 icmp_errors_use_inbound_ifaddr=0 icmp_ignore_bogus_error_responses=1 icmp_ratelimit=1000 icmp_ratemask=6168 igmp_max_memberships=20 igmp_max_msf=10 inet_peer_gc_maxtime=120 inet_peer_gc_mintime=10 inet_peer_maxttl=600 inet_peer_minttl=120 inet_peer_threshold=65664 ip_default_ttl=64 ip_dynaddr=0 ip_forward=0 ipfrag_high_thresh=262144 ipfrag_low_thresh=196608 ipfrag_max_dist=64 ipfrag_secret_interval=600 ipfrag_time=30 ip_local_port_range=32768 61000 ip_local_reserved_ports= ip_nonlocal_bind=0 ip_no_pmtu_disc=0 ping_group_range=1 0 rt_cache_rebuild_count=4 tcp_abc=0 tcp_abort_on_overflow=0 tcp_adv_win_scale=2 tcp_allowed_congestion_control=cubic reno tcp_app_win=31 tcp_available_congestion_control=cubic reno tcp_base_mss=512 tcp_challenge_ack_limit=100 tcp_congestion_control=cubic tcp_dma_copybreak=262144 tcp_dsack=1 tcp_ecn=2 tcp_fack=1 tcp_fin_timeout=60 tcp_frto=2 tcp_frto_response=0 tcp_keepalive_intvl=75 tcp_keepalive_probes=9 tcp_keepalive_time=7200 tcp_limit_output_bytes=131072 tcp_low_latency=0 tcp_max_orphans=262144 tcp_max_ssthresh=0 tcp_max_syn_backlog=2048 tcp_max_tw_buckets=262144 tcp_mem=2316864 3089152 4633728 tcp_min_tso_segs=2 tcp_moderate_rcvbuf=1 tcp_mtu_probing=0 tcp_no_metrics_save=0 tcp_orphan_retries=0 tcp_reordering=3 tcp_retrans_collapse=1 tcp_retries1=3 tcp_retries2=15 tcp_rfc1337=0 tcp_rmem=4096 87380 4194304 tcp_sack=1 tcp_slow_start_after_idle=0 tcp_stdurg=0 tcp_synack_retries=5 tcp_syncookies=1 tcp_syn_retries=5 tcp_thin_dupack=0 tcp_thin_linear_timeouts=0 tcp_timestamps=0 tcp_tso_win_divisor=3 tcp_tw_recycle=0 tcp_tw_reuse=0 tcp_window_scaling=1 tcp_wmem=4096 65536 4194304 tcp_workaround_signed_windows=0 udp_mem=2316864 3089152 4633728 udp_rmem_min=4096 udp_wmem_min=4096 xfrm4_gc_thresh=4194304
2.6.32 is over 5 years old (Dec 2009). I suspect this bug is in the old FRTO code that Yuchung removed when he rewrote FRTO from scratch in 2013. Can you please test a newer version of the kernel (preferably 3.19 but at least 3.12 or newer) to see if this bug is still around? Thanks!
As suggested by David I will repost it in redhat as it seems a RHEL6 component. Thanks!