{{tag>Brouillon Réseau Linux Kernel TCP CA}} = Réseau Linux pile TCP/IP Voir aussi : * MPTCP, SCTP, DCCP Voir : * https://docs.kernel.org/networking/ip-sysctl.html * https://www.inetdoc.net/guides/lartc/lartc.kernel.obscure.html * https://man7.org/linux/man-pages/man7/tcp.7.html * https://frsag.frsag.narkive.com/IYrhTt32/incidence-des-tcp-timestamps-sur-les-connexions * hping2 man 7 tcp == Contrack Voir : * /proc/net/nf_conntrack * /proc/sys/net/nf_conntrack_max apt-get install conntrack Flush conntrack -F === /proc/sys/net/ipv4/tcp_syn_retries $ sysctl net.ipv4.tcp_syn_retries net.ipv4.tcp_syn_retries = 6 Effectively, this takes 1+2+4+8+16+32+64=127s before the connection finally aborts. === /proc/sys/net/ipv4/tcp_synack_retries === /proc/sys/net/ipv4/tcp_retries2 Voir : * https://stackoverflow.com/questions/5227520/how-many-times-will-tcp-retransmit * https://pracucci.com/linux-tcp-rto-min-max-and-tcp-retries2.html Voir aussi : * /proc/sys/net/ipv4/tcp_retries * /proc/sys/net/ipv4/tcp_syn_retries * /proc/sys/net/ipv4/tcp_synack_retries ==== Cluster In a High Availability (HA) situation consider decreasing the setting to 3. RFC 1122 recommends at least 100 seconds for the timeout, which corresponds to a value of at least 8. Oracle suggest a value of 3 for a RAC configuration. Source : https://access.redhat.com/solutions/726753 ==== Nb de retransmissions vs temps An experiment confirms that (on a recent Linux at least) the timeout is more like 13s with the suggested net.ipv4.tcp_retries2=5 “Windows defaults to just 5 retransmissions which corresponds with a timeout of around 6 seconds.” “Five retransmissions corresponds with a timeout of around six seconds.” tcp_retries2=5 means timeout with first transmission plus 5 retransmissions: 12.6 seconds=(2^6 - 1) * 0.2. tcp_retries2=15: 924.6 seconds=(2^10 - 1) * 0.2 + (16 - 10) * 120. Source : https://github.com/elastic/elasticsearch/issues/102788 Voir aussi : https://www.elastic.co/guide/en/elasticsearch/reference/current/system-config-tcpretries.html#_related_configuration === F_RTO https://access.redhat.com/solutions/4978771 == TCP keepalive Voir : * https://blog.cloudflare.com/when-tcp-sockets-refuse-to-die * Python Scripts https://github.com/cloudflare/cloudflare-blog/tree/master/2019-09-tcp-keepalives tcp_keepalive_time https://www.veritas.com/support/en_US/article.100028680 === Configuring TCP/IP keepalive parameters for high availability clients (JDBC) * tcp_keepalive_probes - the number of probes that are sent and unacknowledged before the client considers the connection broken and notifies the application layer * tcp_keepalive_time - the interval between the last data packet sent and the first keepalive probe * tcp_keepalive_intvl - the interval between subsequent keepalive probes * tcp_retries2 - the maximum number of times a packet is retransmitted before giving up echo "6" > /proc/sys/net/ipv4/tcp_keepalive_time echo "1" > /proc/sys/net/ipv4/tcp_keepalive_intvl echo "10" > /proc/sys/net/ipv4/tcp_keepalive_probes echo "3" > /proc/sys/net/ipv4/tcp_retries2 Source : https://www.ibm.com/docs/en/db2/9.7?topic=ctkp-configuring-operating-system-tcpip-keepalive-parameters-high-availability-clients ss -o == Process / diag tools Voir : * https://access.redhat.com/solutions/30453 == Outils === TCP retransmissions Voir : * http://arthurchiao.art/blog/tcp-retransmission-may-be-misleading/ * net.ipv4.tcp_early_retrans Outils : * tcpretrans.bt (bpftrace) * tcpretrans ([[https://github.com/brendangregg/perf-tools/blob/master/net/tcpretrans|perf-tools]]) * tcpretrans.py ([[https://github.com/iovisor/bcc/blob/master/tools/tcpretrans.py|bpfcc-tools - iovisor/bcc]])) Connaitre le rto_min et le rto_max # grep ^Tcp /proc/net/snmp |column -t |cut -c1-99 Tcp: RtoAlgorithm RtoMin RtoMax MaxConn ActiveOpens PassiveOpens AttemptFails EstabResets Tcp: 1 200 120000 -1 6834 964 161 4614 yum install bpftrace /usr/share/bcc/tools/tcpretrans timeout 60 ./tcpretrans | nl sar -n ETCP sar -n TCP # netstat -s |egrep 'segments retransmited|segments send out' 107428604792 segments send out 47511527 segments retransmited # echo "$(( 47511527 * 10000 / 107428604792 ))" 4 https://www.ibm.com/support/pages/tracking-tcp-retransmissions-linux ''tcpretransmits.sh'' #! /usr/bin/bash test -x /usr/sbin/tcpretrans.bt && TCPRETRANS=/usr/sbin/tcpretrans.bt test -x /usr/share/bpftrace/tools/tcpretrans.bt && TCPRETRANS=/usr/share/bpftrace/tools/tcpretrans.bt # https://github.com/brendangregg/perf-tools/blob/master/net/tcpretrans test -x ./tcpretrans.pl && TCPRETRANS=./tcpretrans.pl OUT=/tmp/tcpretransmits.log if [ -z "$TCPRETRANS" ]; then echo "It looks like 'bpftrace' is not installed" else date > $OUT netstat -s |awk '/segments sen. out$/ { R=$1; } /segments retransmit+ed$/ { printf("%.4f\n", ($1/R)*100); }' >> $OUT $TCPRETRANS | tee -a $OUT netstat -s |awk '/segments sen. out$/ { R=$1; } /segments retransmit+ed$/ { printf("%.4f\n", ($1/R)*100); }' >> $OUT fi Resolving The Problem \\ TCP retransmissions are almost exclusively caused by failing network hardware, not applications or middleware. Report the failing IP pairs to a network administrator. == Autres ** horodatages TCP ** https://access.redhat.com/documentation/fr-fr/red_hat_enterprise_linux/9/html/monitoring_and_managing_system_status_and_performance/benefits-of-tcp-timestamps_tuning-the-network-performance tcp_low_latency (Boolean; default: disabled; since Linux 2.4.21/2.6; obsolete since Linux 4.14) net.ipv4.tcp_timestamps = 1 net.ipv4.tcp_window_scaling = 1 net.ipv4.tcp_sack = 1 net.ipv4.tcp_moderate_rcvbuf = 1 ---------------- # ip route get 192.168.100.11 192.168.100.11 dev virbr1 src 192.168.100.1 uid 1000 cache # ip route show dev virbr1 192.168.100.0/24 proto kernel scope link src 192.168.100.1 # ip route change dev virbr1 192.168.100.0/24 proto kernel scope link src 192.168.100.1 rto_min 8ms