Document

Report
Rick McGeer
Chief Scientist, US IGNITE
December 9, 2013
Distributed Clouds and
Software Defined
Networking
Complementary Technologies for
the Next-Generation Internet
Or, A Post-Hoc
Justification for the Last
10 Years of My Life
3
4
The Future is Distributed
Clouds integrated with
Software-DefinedNetworks!
5
SDN is a set of
abstractions over the
networking control
plane
Proxies are an
essential element of
the Internet
Architecture
Shouldn’t
there be an
abstraction
architecture
for proxies?
6
Network Challenges
• Original Concept of the Network: dumb pipe
between smart endpoints
– Content-agnostic routing
– Rates controlled by endpoints
– Content- and user-agnostic forwarding
• Clean separation of concerns
– Routing and forwarding by network elements
– Rate control, admission control, security at
endpoints
Clean separation of
concerns doesn’t work very
well
• Need application-aware stateful forwarding
(e.g., multicast)
• Need QoS guarantees and network-aware
endpoints
– For high-QoS applications
– For lousy links
• Need in-network security and admission
control
– Endpoint security easily overwhelmed…
Some Examples
•
•
•
•
•
•
•
•
•
Load-balanced end-system multicast
Adaptive/DPI-based Intrusion Detection
In-network transcoding to multiple devices
Web and file content distribution networks
Link-sensitive store-and-forward connection-splitting TCP
proxies
Email proxies (e.g., MailShadow)
In-network compression engines (Riverbed)
Adaptive firewall
In-situ computation for data reduction from high-bandwidth
sensors (e.g., high-resolution cameras)
Common Feature
• All of these examples require some combination of in-network and
endpoint services
– Information from the network
– Diversion to a proxy
– Line-rate packet filtering
• All require endpoint processing
– Stateful processing
– Connection-splitting
– Filesystem access
• Three central use cases
– Optimization of network resources, especially bandwidth
– Proximity to user for real-time response
– In-situ sensor processing
Historic Solution:
Middleboxes
• Dedicated network appliances to perform specific function
• Gets the job done, but…
– Appliances proliferate (one or more per task)
– Opaque
– Interact unpredictably…
• Don’t do everything
•
– E.g., generalized in-situ processing engine for data reduction
APST, 2005: “The ability to support…multiple coexisting overlays [of
proxies]…becomes the crucial universal piece of the [network] architecture.”
OpenFlow and SDN
• L2/L3 Technology to permit software-defined control of network
forwarding and routing
• What it’s not:
–
–
–
–
–
On-the-fly software decisions about routing and forwarding
In-network connection-splitting store-and-forward
In-network on-the-fly admission control
In-network content distribution
Magic….
• What it is:
– Table-driven routing and forwarding decisions (including drop and multicast)
– Callback protocol from a switch to a controller when entry not in table (“what do I
do now?”)
– Protocol which permits the controller to update the switch
In-Network Processing
• L4/L7 Services provided by nodes in the network
–
–
–
–
–
TCP/Application layer proxies
Stateful/DPI based intrusion detection
Application-layer admission control
Application-layer load-balancing
….
• Key features
– Stateful processing
– Transport/Application layer information required
Middleboxes and the
Network
• Classic View: Proxies and Middleboxes are a necessary evil
that breaks the “end-to-end principle” (Network should be a
dumb pipe between endpoints)
• Modern View (Peterson): “Proxies play a fundamental role in
the Internet architecture: They bridge discontinuities between
different regions of the Internet. To be effective, however,
proxies need to coordinate and communicate with each
other.”
• Generalized Modern View (this talk): Proxies and
Middleboxes are special cases of a general need: endpoint
processing in the network. We need to merge the Cloud and
the Network.
Going From Today to
Tomorrow
• Today: Middleboxes
• Tomorrow: In-network general-purpose processors fronted by
OpenFlow switches
• Advantages of Middleboxes
– Specialized processing at line rate
• Disadvantages of middleboxes
–
–
–
–
–
Nonexistent programming environment
Opaque configuration
Vendor-specific updates
Only common functions get done
Interact unpredictably…
Anatomy of a Middlebox
Incoming
Packets
ASIC
L2/L3/
DPI
Packet
Filter
Embedded
Linux
Processing
Engine
Packet Output
Generalized Architecture
Incoming
Packets
L2/L3
Packet
Filter
Software
on a
Processing
Engine
Outgoing Packets
The Future
Incoming
Packets
OpenFlow
Switch
Controllers
+
Small
Cloud
Packet Output
Advantages of the
Generalizing and Factoring
the Middlebox
• Transparent
• Open programming environment: Linux +
OpenFlow
• Much broader range of features and functions
• Interactions between middleboxes mediated by
OpenFlow rules
– Verifiable
– Predictable
• Updates are software uploads
OpenFlow + In Network
Processing
+ Line-rate processing
+ Largely implementable on COTS switches
+ Packet handling on a per-flow basis
+ Rapid rule update
+ Unified view of the network
+ L2-L7 services
But I Need Proxies
Everywhere…
• Proxies are needed where I need endpoint processing
– In-situ data reduction
– Next to users
– Where I need filtering
• Can’t always predict these in advance for every service!
• So I need a small cloud everywhere, so I can instantiate a
middlebox anywhere
• Solution = Distributed “EC2” + OpenFlow network
• “Slice”: Virtual Network of Virtual Machines
• OpenFlow creates Virtual Network
• “EC2” lets me instantiate VM’s everywhere
Shenker’s SDN Architecture
Virtual
Network
Network "Operating
System"
OpenFlow
Physical
Network
Specification of a virtual
network, with explicit
forwarding instructions
Translation onto
OpenFlow rules on
physical network
Effectuation on physical
network
26
Perfect for L1-L3
Application
Virtual
Network
Transport
IP
MAC
PHY
Network "Operating
System"
OpenFlow
Physical
Network
27
Key Function we want: Add
Processing Anywhere in the
Virtual Network
Application
Virtual
Distributed
System
Transport
IP
Distributed System
"Operating System"
MAC
OpenFlow + Cloud
Managers
PHY
Physical
Distributed
Cloud
28
Going from Virtual Network
to Virtual Distributed
System
Virtual
Distributed
System
Distributed System
"Operating System"
OpenFlow + Cloud
Managers
Physical
Distributed
Cloud
Specification of a virtual
distributed cloud, with
explicit forwarding
instructions BETWEEN
specified VMs
Translation onto OpenFlow
rules on physical network
AND instantiation on physical
machines at appropriate sites
Effectuation on physical
network AND physical
clouds
29
Key Points
• Federated Clouds can be somewhat heterogeneous
– Must support common API
– Can have some variants (switch variants still present a
common interface through OpenFlow)
• DSOS is simply a mixture of three known components:
– Network Operating System
– Cloud Managers (e.g., ProtoGENI, Eucalytpus,
OpenStack)
– Tools to interface with Network OS and Cloud Managers
(nascent tools under development)
30
Implications for
OpenFlow/SDN
• Southbound API (i.e., OpenFlow): minimal and
anticipated in 1.5
– “Support for L4/L7 services”, aka, seamless redirection
• Northbound API
–
–
–
–
Joint allocation of virtual machines and networks
Location-aware allocation of virtual machines
WAN-aware allocation of networks
QoS controls between sites
• Build on/extend successful architectures
– “Neutron for the WAN”
31
Implications for Cloud
Architectures
• Key problem we’ve rarely considered: how do we easily
instantiate and stitch together services at multiple
sites/multiple providers?
• Multiple sites is easy, multiple providers is not
• Need easy way to instantiate from multiple providers
– Common AUP/Conventions? Probably
– Common form of identity/multiple IDs? Multiple or bottom-up
(e.g. Facebook)
– Common API? Absolutely
• Need to understand what’s important and what isn’t
– E.g. very few web services charge for bandwidth
32
Initial Attempts
•
•
•
•
•
Ignite Technical Architecture/GENI Racks
GENI Mesoscale
SAVI
JGN-X
…
33
With Credit To…
34
GENI Mesoscale
• Nationwide network of small local clouds
• Each cloud
– 80-150 worker cores
– Several TB of disk
– OpenFlow-native local switching
• Interconnected over OpenFlow-based L2 Network
• Local “Aggregate Manager” (aka controller)
• Two main designs with common API
– InstaGENI (ProtoGENI-based)
– ExoGENI (ORCA/OpenStack-based)
• Global Allocation through federate aggregate managers
• User allocation of networks and slices through tools (GENI portal,
Flack)
35
GENI And The Distributed
Cloud Stack
• Physical Resources
– GENI Racks, Emulab, GENI backbone
• Cloud OS
– ProtoGENI, ExoGENI…
• Orchestration Layer
– GENI Portal, Flack…
36
Instageni rack topology
of 222
37
©2010 HP Created on xx/xx/xxxx
U.S. Ignite City Technical Architecture
Existing head-end
Existing
ISP
connects
Most
equipment not
shown
Layer 3 GENI
control plane
Layer 2
Ignite
Connect
(1 GE or
10GE)
Layer 2 connect
to subscribers
OpenFlow switch(es)
Flowvisor
Remote management
Instrumentation
Aggregate manager
Measurement
Programmable servers
Storage
Video switch (opt)
New GENI / Ignite rack pair
Home
GENI Mesoscale Deployment
39
Distributed Clouds and
NSFNet: Back to the Future
•
•
•
•
GENI today is NSFNet circa 1985
GENI and the SFA: Set of standards (e.g., TCP/IP)
Mesoscale: Equivalent to NSF Backbone
GENIRacks: Hardware/software instantiation of
standards that sites can deploy instantly
– Equivalent to VAX 11 running Berkeley Unix
– InstaGENI cluster running ProtoGENI and OpenFlow
• Other instantiations which are interoperable
– VNode (Aki Nakao, University of Tokyo and NICT)
– Tomato (Dennis Schwerdel, TU-Kaiserslautern)
JGN-X (Japan)
41
SAVI (Canada)
42
Ofelia (EU)
43
“Testbeds” vs “Clouds”
• JGN-X, GENI, SAVI, Ofelia, GLab, OneLab
are all described as “Testbeds”
– But they are really Clouds
– Tests require realistic services
• History of testbeds:
– Academic ResearchAcademic/Research
servicesCommercial services
– Expect similar evolution here (but commercial will
come faster)
44
Programming Environment
for Distributed Clouds
• Problem: Allocating and configuring distributed clouds is a
pain
– Allocate network of VM’s
– Build VM’s and deploy images
– Deploy and run software
• But most slices are mostly the same
• Automate commonly-used actions and pre-allocate typical
slices
• 5-minute rule: Build, deploy, and execute “Hello, World” in five
minutes
• Decide what to build: start with sample application
45
TransGeo: A Model
TransCloud Application
• Scalable, Ubiquitous Geographic Information System
• Open and Public
– Anyone can contribute layers
– Anyone can host computation
• Why GIS?
–
–
–
–
–
–
46
Large and active community
Characterized by large data sets (mostly satellite images)
Much open-source easily deployable software, standard data formats
Computation naturally partitions and is loosely-coupled
Collaborations across geographic regions and continents
Very pretty…
TransGeo Architecture
47
TransGeo Sites (May 2013)
48
49
50
Opening up TransGEO: The
GENI Experiment Engine
• Key Idea: Genericize and make available the infrastructure
behind the TransGEO demo
– Open to every GENI, FIRE, JGN-X, Ofelia, SAVI…experimenter
who wants to use it
• TransGEO is a trivial application on a generic infrastructure
– Perhaps 1000 lines of Python code on top of
•
•
•
•
•
•
51
Key-Value Store
Layer 2 network
Sandboxed Python programming environment
Messaging Service
Deployment Service
GIS Libraries
GENI Experiment Engine
•
•
•
•
•
•
•
•
•
52
Permanent, Long-Running, Distributed File System
Permanent, Long-Running, GENI-wide Message Service
Permanent, Long-Running, Distributed Python Environment
Permanent, world-wide Layer-2 VLANs on high-performance
networks
All offered in slices
All shared by many experimenters
Model: Google App Engine
Advantage for GENI: Efficient use of resources
Advantage for Experimenters: Up and running in no time
GENI Experiment Engine Architecture
53
Staged Rollout
• Permanent Layer-2 Network Spring 2014
• Shared File System based on (Swift)
Spring 2014
• Python environment Summer 2014
54
Thanks and Credits
Joe Mambretti, Fei Yeh, Jim Chen
Northwestern/ iCAIR
Andy Bavier, Marco Yuen, Larry Peterson,
Jude Nelson, Tony Mack
PlanetWorks/Princeton
Chris Benninger, Chris Matthews, Chris
University of Victoria
Pearson, Andi Bergen, Paul Demchuk, Yanyan
Zhuang, Ron Desmarais, Stephen Tredger,
Yvonne Coady, Hausi Muller
Heidi Dempsey, Marshall Brinn, Vic Thomas,
Niky Riga, Mark Berman, Chip Elliott
BBN/GPO
Rob Ricci, Leigh Stoller, Gary Wong
University of Utah
Glenn Ricart, William Wallace, Joe Konstan
US Ignite
Paul Muller, Dennis Schwerdel
TU-Kaiserslautern
Amin Vahdat, Alvin AuYoung, Alex Snoeren,
Tom DeFanti
UCSD
55
Thanks and Credits
Nick Bastin
Barnstormer Softworks
Shannon Champion
Matrix Integration
Jessica Blaine, Jack Brassil, Kevin Lai,
Narayan Krishnan, Dejan Milojicic, Norm
Jouppi, Patrick Scaglia, Nicki Watts, Michaela
Mezo, Bill Burns, Larry Singer, Rob Courtney,
Randy Anderson, Sujata Banerjee, Charles
Clark
HP
Aki Nakao
University of Tokyo
56
Conclusions
•
Distributed Clouds are nothing new…
– Akamai was basically the first Distributed Cloud
– Single Application, now generalizing
•
But this is OK…
– Web simply wrapped existing services
•
•
Now in vogue with telcos (“Network Function Virtualization”)
What’s new/different in GENI/JGN-X/SAVI/Ofelia….
– Support from programmable networks
– “Last frontier” for software in systems
•
Open Problems
– Siting VMs!
– Complex network/compute/storage optimization problems
•
Needs
– “http”-like standardization of APIs at IaaS, PaaS layers
57
Links
http://citeseerx.ist.psu.edu/viewdoc/download?d
oi=10.1.1.20.123&rep=rep1&type=pdf
http://pages.cs.wisc.edu/~akella/CS838/F09/838Papers/APST05.pdf
http://www.youtube.com/watch?v=eXsCQdshMr4

similar documents