Reflections on data plane performance, iptables and ipsets · •Calico in a public cloud uses IP-in-IP, with tunnel MTU = 1440 •1440 was optimised for GCE, which has an MTU of

Reflections on data plane performance, iptables and ipsets

Neil Jerram – Metaswitch & Project Calico

@neiljerram www.projectcalico.org

Who am I?

• Free software hacker since 1990s

• Metaswitch (previously Data Connection) since 1995

; line+.el

;

; version 1.1

;

; This has not (yet) been accepted by the Emacs Lisp archive,

; but if it is the archive entry will probably be something like this:

;; line+|Neil Jerram|[email protected]|

;; Line Numbering & Interrupt Driven Actions|

;; 1993-02-18|1.1|<archive pathname of line+.el>|

; Mished and mashed by Neil Jerram <[email protected]>,

; Monday 21 December 1992.

Free software work

• Emacs

• Guile

• Openmoko and GTA04 smartphones

Metaswitch and Project Calico

• 30+ year provider of high quality networking software, but mostly proprietary

• Software -> hardware -> and now back again!

• Now also leading projects as open source• Project Clearwater

• Project Calico

So, Calico?

• Connectivity and security for workloads (aka endpoints, aka micro-services, aka containers or VMs) in an elastic computing environment• e.g. a data center

• Emphasis on simplicity and scalability

• Based on standard Linux features• routing, iptables• and Internet protocols (BGP)

• Mainline case L3 only

Old, zone-based security

Services in an elastic environment

Distributed firewall security

Calico architecture

Data plane performance questions

• Can we get same bandwidth between endpoints as between those endpoints’ hosts?

• What is CPU cost, and how does it compare with other networking approaches?

• What are the effects of our iptables and ipset programming?

Testing methodology

• Two hosts, directly connected by 10Gb link• 8 core

• 64Gb RAM

• 3.13 kernel

• No tuning

• qperf, using TCP

• Measure CPU usage, raw throughput and packet latency

Configurations

• Bare metal, i.e. host to host

• Between OpenStack VMs• ‘TAP’ interface between VM and host

• Between containers• veth pair between container namespace and host namespace

• Between OpenStack VMs using Open vSwitch (OVS) and VXLAN

• MTU 1500, send sizes 20000 and 500

Data plane throughput

• Saturation for 20k messages …(red bars)

• … but not for 500 messages(blue bars)

• Why?• OpenStack better than bare metal?

• OVS case reaches >8Gb/s if MTU is increased to 9000

CPU usage

• CPU-limited for small messages

• OpenStack cases can use more cores

• Extra CPU cost for virtualization• Namespace

• TAP or veth interface

• Routing in guest as well as host

CPU usage per throughput

• CPU required to drive each Gb/s of throughput

Latency

• Tiny extra latency for containers

• More for VMs• But acceptable

• Note micro seconds• Not milli!

Security rules

iptables and ipsets

• iptables on a given host should be the composition of many logical security rules

• Will this impact data plane performance?

• Actually, no

-A felix-FORWARD -i tap+ -j felix-FROM-ENDPOINT-A felix-FORWARD -o tap+ -j felix-TO-ENDPOINT-A felix-FORWARD -i tap+ -j ACCEPT-A felix-FORWARD -o tap+ -j ACCEPT-A felix-FROM-ENDPOINT -i tap7f470881-51 -g felix-from-7f470881-51-A felix-FROM-ENDPOINT -j DROP-A felix-INPUT -i tap+ -j felix-FROM-ENDPOINT-A felix-INPUT -i tap+ -j ACCEPT-A felix-TO-ENDPOINT -o tap7f470881-51 -g felix-to-7f470881-51-A felix-TO-ENDPOINT -j DROP-A felix-from-7f470881-51 -m conntrack --ctstate INVALID -j DROP-A felix-from-7f470881-51 -m conntrack --ctstate RELATED,ESTABLISHED -j RETURN-A felix-from-7f470881-51 -p udp -m udp --sport 68 --dport 67 -j RETURN-A felix-from-7f470881-51 -s 10.28.0.40/32 -m mac --mac-source FA:16:3E:4E:7A:0E -g felix-p-_6b340324948a39b-o-A felix-from-7f470881-51 -m comment --comment "Anti-spoof DROP (endpoint 7f470881-5156-47ce-a67d-b971ef5e5cde):" -j DROP-A felix-p-_6b340324948a39b-i -p icmp -m set --match-set felix-v4-_6b340324948a39b src -j RETURN-A felix-p-_6b340324948a39b-i -s 172.18.203.20/32 -p tcp -m multiport --dports 22 -j RETURN-A felix-p-_6b340324948a39b-i -s 172.18.203.20/32 -p udp -m multiport --dports 5060 -j RETURN-A felix-p-_6b340324948a39b-i -s 172.18.203.20/32 -p tcp -m multiport --dports 80 -j RETURN-A felix-p-_6b340324948a39b-i -m comment --comment "Default DROP rule (72d696a9-f715-495f-9152-7f5e6a69fd0f):" -j DROP

What saves us?

• conntrack

• ipsets scale well, thanks to hash table implementation

• Nested design for source/destination interface mapping

Arjan Schaaf’s measurements

What is happening here?

• http://www.slideshare.net/ArjanSchaaf/docker-network-performance-in-the-public-cloud

• Various approaches to networking between containers on AWS hosts• For this case Calico uses IP-in-IP between the hosts

• Calico bandwidth less than half of native

• We set up the same system, got same results as Arjan• For t2.micro bandwidth = 65.3 MB/sec compared with native = 125 MB/sec.• For m4.xlarge bandwidth = 108 MB/sec compared with native = 267 MB/sec• Why?

It’s all about the MTU

• Calico in a public cloud uses IP-in-IP, with tunnel MTU = 1440

• 1440 was optimised for GCE, which has an MTU of 1460 on its VM interfaces

• But AWS instances have an MTU = 9001!• So native tests were using jumbo frames, and the calico test was using 1440.

• If Calico’s tunnel MTU is increased to 8980• For t2.micro, Calico bandwidth = 114 MB/sec

• For m4.xlarge, Calico bandwidth = 266 MB/sec

• Problem solved – Calico throughput is now close to native

So what have we learned?

• With Calico connectivity, VMs or containers can saturate a 10Gb link between hosts, just as much as the hosts themselves could

• There is a CPU cost to virtualization• But mostly inevitable if you want virtualization at all (non-accelerated)

• Calico does not add any significant extra cost

• Conntrack largely saves us from the effects of complex iptables• ipsets and clever programming design also help

• Be humble about performance comparisons

Further information, and thanks!

• Project Calico• http://www.projectcalico.org/• http://docs.projectcalico.org/en/latest/• https://github.com/projectcalico

• Blog on Calico data plane performance• http://www.projectcalico.org/calico-dataplane-performance/

• Thanks!• @neiljerram• @projectcalico• www.metaswitch.com