| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <1410875959.7106.200.camel@edumazet-glaptop2.roam.corp.google.com> Date: Tue, 16 Sep 2014 06:59:19 -0700 From: Eric Dumazet <eric.dumazet@...il.com> To: Jesper Dangaard Brouer <brouer@...hat.com> Cc: "netdev@...r.kernel.org" <netdev@...r.kernel.org>, Stephen Hemminger <stephen@...workplumber.org>, Tom Herbert <therbert@...gle.com>, David Miller <davem@...emloft.net>, Hannes Frederic Sowa <hannes@...essinduktion.org>, Daniel Borkmann <dborkman@...hat.com>, Florian Westphal <fw@...len.de>, Toke Høiland-Jørgensen <toke@...e.dk>, Dave Taht <dave.taht@...il.com> Subject: Re: Qdisc: Measuring Head-of-Line blocking with netperf-wrapper On Tue, 2014-09-16 at 15:22 +0200, Jesper Dangaard Brouer wrote: > Zooming in on high-prio ping only, and comparing TSO vs GSO: > http://people.netfilter.org/hawk/qdisc/measure01/compare_TSO_vs_GSO__ping_hiprio.png > http://people.netfilter.org/hawk/qdisc/measure01/compare_TSO_vs_GSO__ping_cdf.png > > - It clearly shows that GSO have lower/better ping values that TSO, > e.g. smaller HoL blocking If you use a single TX queue on the NIC (no prio aware), this makes no sense.... GSO is exactly sending the same packet train on the wire, it only uses more cpu on the host to segment packets. While doing the segmentation, no packet can be sent to the NIC, so this should be _adding_ some latencies, unless the TSO engine on your NIC is brain damaged. The high prio ping packet cannot be inserted in the middle of a GSO train. ping is not very good to measure very small rtt actually. With the TCP usec rtt work I did lately, you'll get more precise results from a TCP_RR flow, as Tom and I explained. Proof : lpaa23:~# ping -n -c 10 10.246.7.152 PING 10.246.7.152 (10.246.7.152) 56(84) bytes of data. 64 bytes from 10.246.7.152: icmp_req=1 ttl=64 time=0.196 ms 64 bytes from 10.246.7.152: icmp_req=2 ttl=64 time=0.161 ms 64 bytes from 10.246.7.152: icmp_req=3 ttl=64 time=0.183 ms 64 bytes from 10.246.7.152: icmp_req=4 ttl=64 time=0.070 ms 64 bytes from 10.246.7.152: icmp_req=5 ttl=64 time=0.195 ms 64 bytes from 10.246.7.152: icmp_req=6 ttl=64 time=0.163 ms 64 bytes from 10.246.7.152: icmp_req=7 ttl=64 time=0.169 ms 64 bytes from 10.246.7.152: icmp_req=8 ttl=64 time=0.183 ms 64 bytes from 10.246.7.152: icmp_req=9 ttl=64 time=0.150 ms 64 bytes from 10.246.7.152: icmp_req=10 ttl=64 time=0.208 ms --- 10.246.7.152 ping statistics --- 10 packets transmitted, 10 received, 0% packet loss, time 8999ms rtt min/avg/max/mdev = 0.070/0.167/0.208/0.040 ms While with TCP_RR, you'll get the ~20 usec rtt we really have on this link : netperf -H 10.246.7.152 -t TCP_RR -l 100 & ss -temoi dst 10.246.7.152 ... ESTAB 0 1 10.246.7.151:59804 10.246.7.152:45623 timer:(on,201ms,0) ino:14821 sk:ffff881fcb48a740 <-> skmem:(r0,rb357120,t2304,tb46080,f1792,w2304,o0,bl0) ts sack cubic wscale:6,6 rto:201 rtt:0.019/0.003 ato:40 mss:1448 cwnd:10 send 6096.8Mbps pacing_rate 11656.9Mbps unacked:1 rcv_rtt:572 rcv_space:29470 And if you take a look at tcpdump, you'll see that 20 usec is the reality. lpaa23:~# tcpdump -p -n -s 128 -i eth0 -c 10 host 10.246.7.152 tcpdump: verbose output suppressed, use -v or -vv for full protocol decode listening on eth0, link-type EN10MB (Ethernet), capture size 128 bytes 06:53:59.278218 IP 10.246.7.152.43699 > 10.246.7.151.49454: P 1088080991:1088080992(1) ack 811259668 win 453 <nop,nop,timestamp 1151215 1152060> 06:53:59.278260 IP 10.246.7.151.49454 > 10.246.7.152.43699: P 1:2(1) ack 1 win 457 <nop,nop,timestamp 1152060 1151215> 06:53:59.278268 IP 10.246.7.152.43699 > 10.246.7.151.49454: P 1:2(1) ack 2 win 453 <nop,nop,timestamp 1151215 1152060> 06:53:59.278220 IP 10.246.7.151.49454 > 10.246.7.152.43699: P 2:3(1) ack 2 win 457 <nop,nop,timestamp 1152060 1151215> 06:53:59.278232 IP 10.246.7.152.43699 > 10.246.7.151.49454: P 2:3(1) ack 3 win 453 <nop,nop,timestamp 1151215 1152060> 06:53:59.278240 IP 10.246.7.151.49454 > 10.246.7.152.43699: P 3:4(1) ack 3 win 457 <nop,nop,timestamp 1152060 1151215> 06:53:59.278252 IP 10.246.7.152.43699 > 10.246.7.151.49454: P 3:4(1) ack 4 win 453 <nop,nop,timestamp 1151215 1152060> 06:53:59.278259 IP 10.246.7.151.49454 > 10.246.7.152.43699: P 4:5(1) ack 4 win 457 <nop,nop,timestamp 1152060 1151215> 06:53:59.278282 IP 10.246.7.152.43699 > 10.246.7.151.49454: P 4:5(1) ack 5 win 453 <nop,nop,timestamp 1151215 1152060> 06:53:59.278289 IP 10.246.7.151.49454 > 10.246.7.152.43699: P 5:6(1) ack 5 win 457 <nop,nop,timestamp 1152060 1151215> ping is slower : lpaa23:~# tcpdump -p -n -s 128 -i eth0 -c 10 host 10.246.7.152 tcpdump: verbose output suppressed, use -v or -vv for full protocol decode listening on eth0, link-type EN10MB (Ethernet), capture size 128 bytes 06:57:29.666168 IP 10.246.7.151 > 10.246.7.152: ICMP echo request, id 11062, seq 4, length 64 06:57:29.666355 IP 10.246.7.152 > 10.246.7.151: ICMP echo reply, id 11062, seq 4, length 64 06:57:30.666220 IP 10.246.7.151 > 10.246.7.152: ICMP echo request, id 11062, seq 5, length 64 06:57:30.666408 IP 10.246.7.152 > 10.246.7.151: ICMP echo reply, id 11062, seq 5, length 64 06:57:31.666147 IP 10.246.7.151 > 10.246.7.152: ICMP echo request, id 11062, seq 6, length 64 06:57:31.666333 IP 10.246.7.152 > 10.246.7.151: ICMP echo reply, id 11062, seq 6, length 64 06:57:32.666164 IP 10.246.7.151 > 10.246.7.152: ICMP echo request, id 11062, seq 7, length 64 06:57:32.666359 IP 10.246.7.152 > 10.246.7.151: ICMP echo reply, id 11062, seq 7, length 64 Really, do not rely too much on ping. If you do, we cant really trust your results. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@...r.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists