DTIC ADA579328: Language Abstractions for Software-Defined Networks

0.37 MB·English

by  Defense Technical Information Center

Language Abstractions for Software-Defined Networks
NateFoster MichaelJ.Freedman RobHarrison ChristopherMonsanto
CornellUniversity PrincetonUniversity USMilitaryAcademy PrincetonUniversity
MarkReitblatt JenniferRexford AlecStory DavidWalker
CornellUniversity PrincetonUniversity CornellUniversity PrincetonUniversity
1. Introduction
Traditional Network
For the past 30 years, networks have been built the same
way: out of special-purpose devices running distributed al- Switch
Control Plane
gorithms that provide functionality such as topology dis-
. Data Plane
covery, routing, traffic monitoring, and access control. Re-
cent years, however, have seen growing interest in a new
kindofnetworkarchitectureinwhichalogically-centralized
controller machine manages a distributed collection of
programmable switches. These software-defined networks .
Software-Defined Network
(SDNs) make it possible for programmers to directly con-
trol the behavior of the network by configuring the packet-
Controller
forwarding rules installed on switches [6]. Figure 1 depicts Programmable Machine
Switch .
the architecture of a traditional network, where each inde-
pendentswitchconsistsoftightly-integratedcontrolanddata
planes, and of an SDN, where a collection of switches are
managedbyasingleprogramrunningonthecontroller.1
SDNs both simplify existing applications and also pro- .
videaplatformfordevelopingexcitingnewones.Forexam- Figure1. Networkarchitectures.
ple, to implement shortest-path routing, the controller can
calculate the forwarding rules for each switch by running
thebehaviorofanetworkatasuitablyhigh-levelofabstrac-
Dijkstra’s algorithm on the graph of the network topology
tion, leaving tedious implementation details to a compiler
rather than running a more complicated distributed proto-
and run-time system. This paper summarizes the results of
col [9]. To enforce a fine-grained access control policy, the
our initial efforts, focusing on three areas where language-
controller can consult an external authentication server and
basedabstractionscanbeleveragedtoprovidetangibleben-
install a custom forwarding path for each user [2]. And to
efitstoprogrammers.
balancetheloadbetweenback-endserversinadatacenter,
the controller can migrate flows in response to server load
2. NetworkPolicyAbstractions
andnetworkcongestion[10].
AlthoughSDNsprovideanarchitecturethatmakesitpos- Today’s SDN controllers support an event-driven program-
sibletoprogramacollectionofswitches,thenetworkisstill mingmodelinwhichprogramsreacttonetworkevents(e.g.,
adistributedsystem,andalloftheusualcomplicationsarise: topologychanges,packetsforwhichaswitchdoesnothave
control messages may be delayed or lost, packets may be a forwarding rule, etc.) by manipulating the rules installed
dropped, and the devices may experience unexpected fail- onswitches,orbycraftingpacketstoemitontothenetwork.
ures.Unfortunately,currentSDNplatformssuchasNOX[4] Thiscomplicatesprogramsinseveralways.Foronething,it
and Beacon [1] provide a low-level interface that supports splitscontrolbetweentwolevelsofexecution—theprogram
manipulating the state of individual devices, but little else. on the controller and the rules on the switches. Program-
Hence,writingapplicationsforSDNstodayisatediousex- mers must constantly consider whether installing or unin-
erciseinlow-leveldistributedprogramming. stalling rules will mask future events received at the con-
ThegoaloftheFreneticprojectistodevelopdeclarative troller, and they must explicitly coordinate multiple asyn-
constructsthatmakeitpossibleforprogrammerstodescribe chronous events to perform simple tasks (e.g., discovering
the switches connected to a particular link). For another, it
1Thecontrollermaybereplicatedforscalabilityandfaulttolerance[5]. forces programmers to manage a host of low-level details
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
JAN 2012 2. REPORT TYPE 00-00-2012 to 00-00-2012
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Language Abstractions for Software-Defined Networks
5b. GRANT NUMBER
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
Princeton University,Department of Computer REPORT NUMBER
Science,Princeton,NJ,08544
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
Proc. Workshop on Languages for Distributed Algorithms (LADA ’12) Philadelphia, PA, January 2012.
U.S. Government or Federal Rights License
14. ABSTRACT

15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON
a. REPORT b. ABSTRACT c. THIS PAGE Same as 3
unclassified unclassified unclassified Report (SAR)
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
involvingtheswitchhardware(e.g.,whethertousewildcard 4. ConsistentUpdateAbstractions
or exact-match rules to implement the desired forwarding
Our most recent work on Frenetic focuses on the problem
policy, and the values to use for rule timeouts and priority
ofimplementingupdatestonetworkpolicy.ManySDNpro-
levels). Frenetic’s NetCore policy language abstracts away
grams need to transition from one policy to another (e.g.,
from many of these details involving the underlying dis-
due to topology changes, planned maintenance, or unex-
tributed system [7]. Instead of having to write programs in
pected failures). From the perspective of the programmer,
termsoflow-levelnetworkeventsandforwardingrules,Net-
it would be ideal if such transitions could be propagated
Coreprogrammerswriteasimplespecificationthatcaptures
atomically to the switches in the network. But atomic up-
theintendedforwardingbehaviorofthenetwork.Thecom-
datesareimpracticaltoimplement—theywouldrequiredis-
pilerpartitionsthisspecificationintosuitablecodefragments
rupting the operation of the network during the update. We
forthecontrollerandtheswitches.Italsogeneratesthecom-
have been developing consistency abstractions that allow
municationpatternsbetweenthecontrollerandswitches,us-
programmers to specify how updates to policy should be
ingarun-timesystemthathandlesthebookkeepingrelated
propagatedtothedevicesinthenetwork[8].Theseabstrac-
toinstallinganduninstallingindividualforwardingrules.A
tions are strong enough to provide useful guarantees about
notablefeatureofNetCoreisthatitallowspoliciestobede-
applicationbehaviors,andyetrelaxedenoughtoadmitprac-
scribedintermsofarbitraryfunctionsthatcannotbedirectly
tical implementations. For example, per-packet consistency
realized on switches. To handle such policies, the compiler
ensuresthateverypacketflowingthroughthenetworkuses
generates an underapproximation of the overall policy us-
a single version of the policy and not a mixture of old and
ingasimplestaticanalysis,andthenusespartialevaluation
new policies. This preserves all properties that can be ex-
to refine the underapproximation at run time using the ac-
pressed in terms of individual packets and the paths they
tualpacketsseeninthenetwork.Lastly,NetCorehasaclear
take through the network—a class of properties that sub-
formal semantics that provides a basis for reasoning about
sumesimportantstructuralinvariantssuchasbasicconnec-
programs.
tivity and loop-freedom, as well as access control policies.
Going a step further, per-flow consistency ensures that sets
3. NetworkQueryAbstractions of related packets are processed with the same policy. This
canbeusedtoenforcerichpropertiesincludingtheordering
AnothercomplicationinSDNprogramstodayconcernsthe
of packets within a flow. It is also needed for applications
implementation of monitoring policies. Each switch main-
suchasloadbalancers,whichneedtoensurethatpacketsin
tains counters for each forwarding rule that keeps track of
thesameflowreachthesamedestinationtoavoidbreaking
the total number of packets and bytes processed using it.
connections.Freneticprovidesanidealplatformforexplor-
To monitor traffic, the controller can poll the counters as-
ing such abstractions, as the compiler and run-time system
sociated with particular rules. However, programmers must
can be used to perform the tedious bookkeeping related to
ensure that the rules installed on switches are fine-grained
implementingnetworkupdates.Forexample,toprovideper-
enough to collect the desired traffic statistics. For instance,
packet consistency, the run-time system can generate rules
to monitor the total amount of web traffic, the program-
that tag packets with policy-version numbers and use two-
mermustinstallrulesthatprocess(andcount)trafficinvolv-
phasecommittopropagaterulestoswitches.Per-flowcon-
ingTCPport80separatelyfromallothertraffic.Managing
sistency can be implemented in similar fashion, with some
thesedetailsistediousandmakesprogramsanti-modular—
additional analysis to identify existing flows. We are cur-
the rules generated by one module may be too coarse to
rentlyworkingtoextendourcompilerandrun-timesystem
be executed side-by-side with a different module. To sup-
withtheseabstractions.
port applications whose correct operation involves a moni-
toringcomponent,Freneticincludesanembeddedquerylan-
5. ExampleFreneticApplication
guage that provides effective abstractions for reading net-
work state [3]. This language, which has syntax reminis- Figure 2 presents code for a Frenetic application that: (1)
centofSQL,includesconstructsforselecting,filtering,split- learns the locations of hosts on the network, (2) forwards
ting,merging,andaggregatingthestreamsofpacketsflow- packets to known destinations, (3) floods packets to un-
ing through the network. Importantly, the language allows known destinations, and (4) periodically reports the largest
queries to be composed with each other and with forward- k sourcesoftraffic.Spaceconstraintsprecludegivingafull
ing policies, without any harmful interference. Again, the explanationofthecode.Detailscanbefoundintheoriginal
technologythatmakesthispossibleisthecompilerandrun- paper on Frenetic [3]. Note however that the program does
timesystem,whichgeneratesswitch-levelrulesguaranteed not include any code that explicitly manipulates low-level
to correctly implement all queries and the overall forward- switchstate,sendscontrolmessagestoswitches,orhandles
ingpolicy.Thecompileralsogeneratesthecontrolmessages asynchronous network events. Instead, the entire program
and handlers needed to query and tabulate the counters on consists of a few declarative specifications that are com-
switches. posedtogetherinasimpleandelegantway.
##################### #######################
# Ethernet learning # # Top-k heavy hitters #
##################### #######################
# Store learned mac addresses in table # Constants
def learn_mac(((switch,mac),packet),table): K = 2
table[switch][mac] = inport(header(packet)) WINDOW = 15
return table
# Store new values in stats
# Convert table to policy def tabulate_stats(new,stats):
def make_policy(table): for host,bytes in new.items():
policy = defaultdict(lambda:[]) stats[host] += int(bytes)
for switch,learned in table.items(): return stats
patterns = false_fp()
for mac,port in learned.items(): # Convert stats to list of top-k hosts
rule = Rule(dstmac_fp(mac),[forward(port)]) def topK(stats):
policy[switch].append(rule) l = sorted(stats.items(), key=lambda x:x[1])
patterns = patterns | dstmac_fp(mac) l.reverse()
flood_rule = Rule(true_fp() - patterns,[flood()]) return l[:K]
policy[switch].append(flood_rule)
return policy # Initial stats
init_stats = defaultdict(lambda:0)
# Initial table
init_table = defaultdict(lambda:{}) # Main function
def heavy_hitters():
# Main function q = Select(’sizes’) * \
def ethernet_learning(): GroupBy([’srcmac’]) * \
q = Select(’packets’) * \ Every(WINDOW)
GroupBy([’switch’,’srcmac’]) * \ sf = Accum(init_stats, tabulate_stats) >> \
SplitWhen([’inport’]) * \ Lift(topK)
Limit(1) return q >> sf >> Print()
sf = Accum(init_table,learn_mac) >> \
Lift(make_policy)
return q >> sf >> Register()
Figure2. Exampleapplication:Ethernetlearningandtop-kheavyhitters.
Acknowledgments OurworkissupportedinpartbyONR tionnetworks. InUSENIXSymposiumonOperatingSystems
grantN00014-09-1-0770andNSFgrantsCNS-1111698and Design and Implementation (OSDI), Vancouver, BC, pages
CNS-1111520. Any opinions, findings, and recommenda- 351–364,October2010.
tions are those of the authors and do not necessarily reflect [6] Nick McKeown, Tom Anderson, Hari Balakrishnan, Guru
theviewsoftheNSF,ONR,ortheUSMilitaryAcademy. Parulkar,LarryPeterson,JenniferRexford,ScottShenker,and
JonathanTurner. Openflow:Enabling innovationin campus
References networks. SIGCOMMCCR,38(2):69–74,2008.
[1] Beacon: A java-based OpenFlow control platform. See [7] ChristopherMonsanto,NateFoster,RobHarrison,andDavid
http://www.beaconcontroller.net,December2011. Walker. A compiler and run-time system for network pro-
gramming languages. In ACM SIGPLAN–SIGACT Sym-
[2] Martin Casado, Michael J. Freedman, Justin Pettit, Jianying
posium on Principles of Programming Languages (POPL),
Luo,NatashaGude,NickMcKeown,andScottShenker. Re-
Philadelphia,PA,January2012. Toappear.
thinkingenterprisenetworkcontrol. IEEE/ACMTransactions
onNetworking,17(4),August2009. [8] Mark Reitblatt, Nate Foster, Jennifer Rexford, and David
Walker. Consistent updates for software-defined networks:
[3] NateFoster,RobHarrison,MichaelJ.Freedman,Christopher
Changeyoucanbelievein!InACMWorkshoponACMWork-
Monsanto, Jennifer Rexford, Alec Story, and David Walker.
shop on Hot Topics in Networks Programmable Routers for
Frenetic: A network programming language. In ACM SIG-
ExtensibleServicesofTomorrow(HotNets),Cambridge,MA,
PLANInternationalConferenceonFunctionalProgramming
November2011.
(ICFP),Tokyo,Japan,September2011.
[9] Scott Shenker, Martin Casado, Teemu Koponen, and Nick
[4] Natasha Gude, Teemu Koponen, Justin Pettit, Ben Pfaff,
McKeown.Thefutureofnetworkingandthepastofprotocols,
Mart´ın Casado, Nick McKeown, and Scott Shenker. NOX:
October2011. InvitedtalkatOpenNetworkingSummit.
Towardsanoperatingsystemfornetworks. SIGCOMMCCR,
38(3),2008. [10] Richard Wang, Dana Butnariu, and Jennifer Rexford.
OpenFlow-basedserverloadbalancinggonewild. InWork-
[5] TeemuKoponen,MartinCasado,NatashaGude,JeremyStri-
shop on Hot Topics in Management of Internet, Cloud, and
bling,LeonPoutievski,MinZhu,RajivRamanathan,Yuichiro
Enterprise Networks and Services (Hot-ICE), Boston, MA,
Iwata, Hiroaki Inoue, Takayuki Hama, and Scott Shenker.
Onix: A distributed control platform for large-scale produc- March2011.