(pdf)

.
Frenetic: Functional Reactive
Programming for Networks
Nate Foster (Cornell)
Mike Freedman (Princeton)
Rob Harrison (Princeton)
Matthew Meola (Princeton)
Jennifer Rexford (Princeton)
David Walker (Princeton)
.
.
.
IBM PLDay 2010
.
Photo credit: http://www. ickr.com/photos/adrianblack
Why Programmable Networks?
Security
• Access control
• Traffic isolation
Monitoring
• Usage / billing
• Anomaly detection
Features
• Virtual Private Networks
• Content Distribution
• Resource Indirection
• Anycast
.
.
3
Current State of Play
It’s a mess!
[Caldwell et al. ’03, Oppenheimer et al. ’03]
4
Current State of Play
It’s a mess!
[Caldwell et al. ’03, Oppenheimer et al. ’03]
Con guration is vendor speci c and complicated
Hodgepodge of mechanisms:
• OSPF / BGP for routing
• ACLs for security
• Net ow for monitoring
Operator errors common and costly
• Outages
• Degraded performance
• Security vulnerabilities
Con guration checkers and lint-like tools help a bit... but
they are only a “band-aid”, not a robust solution
4
This Talk
1. OpenFlow
2. Examples
3. Frenetic
4. Implementation
5. Current and Ongoing work
5
OpenFlow
Traditional Switch
.
.
Control Plane
• General-purpose hardware
• Runs (distributed) routing protocols
• Manipulates the forwarding table in the
data plane
Data Plane
•
Special-purpose hardware
• Implements high-speed forwarding table
• Processes packets at line speed
7
OpenFlow
Key Ideas
• Move control from switch to a stock machine
• Standardize interface between switches and controller
Controller
.
Switches
.
http://www.openflowswitch.org/
8
OpenFlow Switch
Switches process packets using rules described by:
• pattern – identify a set of packets
• priority – disambiguate rules with overlapping patterns
• actions – specify processing of packets
• counters – track number and size of packets processed
9
OpenFlow Switch
Switches process packets using rules described by:
• pattern – identify a set of packets
• priority – disambiguate rules with overlapping patterns
• actions – specify processing of packets
• counters – track number and size of packets processed
Example (OpenFlow Rules)
Pattern
Priority Actions
Counters
[ (OFPAT OUTPUT, PORT 1)
{in port=2, trans src=80} HIGH
(3,1455)
(OFPAT OUTPUT, CONTROLLER) ]
{in port=2}
LOW
[ (OFPAT OUTPUT, PORT 1) ]
(20,12480)
9
OpenFlow Controller
Controller runs a program that responds to events in the
network by installing / uninstalling rules and collecting
statistics from counters.
Event Handlers
• switch join(switch)
• switch leave(switch)
• packet in(switch, inport, packet)
• stats in(switch, pattern, stats)
Messages
• install(switch, pattern, priority, action)
• uninstall(switch, pattern)
• query stats(switch, pattern)
10
Examples
Topology
Controller
1
.
2
Switch
.
12
Static Forwarding
def static forwarding():
# patterns
p1 = {IN PORT:1}
p2 = {IN PORT:2}
.
. #a1actions
= [(OFPAT OUTPUT, PORT 2)]
a2 = [(OFPAT OUTPUT, PORT 1)]
# install rules
install(switch, p1, HIGH, a1)
install(switch, p2, HIGH, a2)
Controller
1
.
2
Switch
.
13
Forwarding + Per-Host Monitoring
def static forwarding per host monitoring():
# patterns
p1 = {IN PORT:1}
p2 = {IN PORT:2}
.
. #a1actions
= [(OFPAT OUTPUT, PORT 2)]
a2 = [(OFPAT OUTPUT, CONTROLLER)]
# install rules
install(switch, p1, HIGH, a2)
install(switch, p2, LOW, a2)
Controller
1
.
2
Switch
.
14
Forwarding + Per-Host Monitoring
defpacket
static forwarding
per host
monitoring():
def
in(switch, inport,
packet):
#
patterns
# patterns
{INSRC:dstmac(packet)
PORT:1}
pp1
= {=DL
}
p2
=
{
IN
PORT:2
}
pweb = {DL
DST:dstmac(packet),
DL TYPE:IP,
. TP SRC:80}
NW PROTO:TCP,
. #a1actions
= [(OFPAT OUTPUT, PORT 2)]
action
. CONTROLLER)]
. #a a2
= [(OFPAT
OUTPUT,
= [(OFPAT
OUTPUT,
PORT 1)]
installrules
rules
# #install
install(switch,pweb,
p1, HIGH,
a2)
install(switch,
HIGH,
a)
install(switch,p,p2,
LOW, a2)
install(switch,
MEDIUM,
a)
Controller
1
.
2
Switch
.
# query counters
query stats(switch, pweb)
14
OpenFlow Limitations
Low-level interface to switch hardware
• priorities used to disambiguate overlapping rules
• no support for negation
• wildcard vs. exact-match rules
Two-tier programming model
• controller program manipulates rules
• asynchronous callbacks
• tricky race conditions
Program pieces don’t compose
• many programs decompose naturally into modules—e.g.,
forwarding + monitoring + access control
• but difficult to program in a compositional style because in general
the rules manipulated by each module will overlap
15
Frenetic
Frenetic Ingredients
High-level pattern algebra
• Hides details of how rules are implemented on switches
• Includes standard logical operators (e.g., negation)
Uni ed programming model
• Programs “see every packet”
• Based on FRP → no asynchronous callbacks
Fully compositional
• Programs can operate on overlapping subsets of the traffic
• Run-time system handles switch-level implementation details
17
Frenetic Ingredients
High-level pattern algebra
• Hides details of how rules are implemented on switches
• Includes standard logical operators (e.g., negation)
Uni ed programming model
• Programs “see every packet”
• Based on FRP → no asynchronous callbacks
Fully compositional
• Programs can operate on overlapping subsets of the traffic
• Run-time system handles switch-level implementation details
Main Challenge: having all these features without
sacri cing performance.
17
Frenetic Core
. Eα
EF α β
Packets
Seconds
Apply
Lift
|O|
First
.
Merge
LoopPre
Calm
Filter
Group
Regroup
Ungroup
event stream carrying values .of type α
operator that transforms an E α into an E β
∈
∈
∈
∈
∈
∈
∈
∈
∈
∈
∈
∈
∈
E packet
E int
(EF a b × E a) → E b
(a → b) → EF a b
EF a b → EF b c → EF a c
EF a b → EF (a × c) (b × c)
.
(E a × E b) → E (a option
× b option)
(c × EF (a × c) (b × c)) → EF a b
EF a a
(a → bool) → EF a a
(a → b) → EF a (b × E a)
((a × a) → bool) → EF (b × E a) (b × E a)
int option × (b × a → b) → b → EF (c × E a) (c × b)
18
Forwarding + Per-Host Monitoring
# sum sizes: (packet list) -> int
def sum sizes(l):
return (reduce(lambda n,p:n + size(p),l,0))
# per host monitoring ef: EF packet (mac * int)
def per host monitoring ef():
return (Filter(inport fp(2) & srcport fp(80)) |O|
Group(dstmac gp()) |O|
ReGroupByTime(30) |O|
Lift(lambda (m,l):(m,sum sizes(l))))
.
.
# E packet
# E (mac * E packet)
# E (mac * packet list)
# E (mac * int)
# rules: (rule list)
rules = [Rule(inport fp(1), [output(2)]),
Rule(inport fp(2), [output(1)])]
# main function
def per host monitoring():
register static(rules)
stats = Apply(Packets(), per host monitoring ef())
print stream(stats)
19
Ethernet Learning
# add rule: (mac * packet) * ((mac * rule) list) -> ((mac * rule) list) * ((mac * rule) list)
def add rule(((m,p),t)): . . .
# complete rules: ((mac * rule) list) -> (rule list)
def complete rules(t): . . .
# learning switch ef: EF packet
def learning switch ef():
return (Group(srcmac gp()) |O|
.
.
Regroup(inport rf()) |O|
Ungroup(1, lambda n,p:p, None) |O|
LoopPre({}, Lift(add rule)) |O|
Lift(complete rules))
# E (mac * E packet)
# E (mac * E packet)
# E (mac * packet)
# E ((mac * rule) list)
# E (rule list)
# main function
def learning switch():
rules = Apply(Packets(), learning switch ef())
register stream(rules)
20
Per-Host Monitoring + Learning
def per host monitoring learning switch():
# ethernet learning
rules = Apply(Packets(), learning switch ef())
.
register stream(rules) .
# per-host monitoring
stats = Apply(Packets(), per host monitoring ef())
print stream(stats)
21
Implementation
Frenetic Program
subscribe
register
Packets
Run-Time System
install
uninstall
.
packet_in
NOX
.
OpenFlow Switches
22
Implementation
Push-based FRP implementation
• Classic pull-based strategy is not a good t for networks
• Frenetic implementation based on strategy developed in FrTime
[Cooper and Krishnamurthi ’06]
Subscribe / Register Library
•
•
•
•
•
Programs can subscribe to streams of packets, headers, ints
They can also register packet-forwarding policies
Semantics is fully compositional
Run-time system manages switch-level rules, event handlers, etc.
Two strategies: proactive (eager) and reactive (lazy)
23
Current and Ongoing Work
Surface Language
• Current prototype is implemented as a Python library
• We want a front end with convenient syntax, typechecker, etc.
Algebraic Optimizer
• Key optimization is moving processing from controller to switches
• Currently programmers must transform programs by hand
• We want an optimizer that rewrites programs automatically
Formal Semantics
• Want a framework for modeling network behavior
• Use to prove optimizations correct
• And to develop new constructs for manipulating traffic atomically
Applications
• Application-level load balancing
• Isolation in multi-tenant networks
24
Questions?
.
Collaborators
Mike Freedman, Rob Harrison, Matt Meola,
Jen Rexford, Dave Walker
http://www.cs.cornell.edu/~jnfoster
25
Similar pages