* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Adopting Ideas from Interplanetary Networking for Sensor
Survey
Document related concepts
Transcript
Adopting Ideas from Interplanetary Networking for Sensor Network Applications Andrew Parker, UCLA Scott Burleigh, JPL Richard Guy, UCLA Deborah Estrin, UCLA 11/12/04 CENS Seminar 1 Interplanetary Network Vision • Picture of sensor networks in space QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. * V. Cerf, InterPlanetary Internet, 2004 11/12/04 CENS Seminar 2 Why We Care 11/12/04 CENS Seminar 3 Outline • • • • • • 11/12/04 Mexico Seismic Array Deployment: Set Up Interplanetary Networking: Set Up & Relevance Delay Tolerant Networking (DTN): Architecture DTN vs. Email Mexico Deployment: DTN Approach Modular DTN Architecture for Sensor Networks CENS Seminar 4 Mexico Deployment • People: Paul Davis, Deborah Estrin, Richard Guy, Martin Lukac, John Wallace, Monica Kohler, Ramesh Govindan, Igor Stubailo, Allen Husker, Katie Mika, John Propst, Sam Irvine, Jeremy Elson, et al. • • • • Seismic array of 50 nodes 5 km apart Spanning 250 km Through jungle, mountain, and urban environments • Targeting end of Q1 2005 11/12/04 CENS Seminar 5 Data Story • CENS Data Communications Controller • Q330 24 bit Digitizer Guralp Seismometer • • – Satellite uplink (expensive) – Data mule (graduate student) Stargate • 11/12/04 CENS Seminar 6 Every bit is important. Can’t lose data. 24 bits / sample, 100 samples / sec, 3 channels = 900 bytes/sec, 3.24 MB/hour, 1.2-1.8 MB/s compressed, per Node 1 GB Flash will hold 20 days of compressed data Conventional methods of retrieval Is there a better way? A Better Way: Networking the Array • Goal: Multihop data from the edge into Mexico City • Mexico deployment is different from typical sensor network applications – Every bit of data is important – Radio connectivity is directional • Due to distance, power, and regulations, we’re using directional antennas – Routing is over a linear topology – Radio communication is not the greatest power consumer 11/12/04 CENS Seminar 7 Braided String • • Introduces some path redundancy 802.11b – 200 mw radio (SMC PCMCIA card) • Yagi antenna – – – – – • 20 dbi 30 degree spread Tested with 2-way splitter 5 - 10 km in LA area 4 Mb/s - 150 kb/s Parabolic antenna – – – – – 11/12/04 24 dbi 5 - 10 degree spread 4 way splitter Tested to ~ 25 km Closer to 1 Mb/s CENS Seminar 8 Routing is Hard to Do Right • Possible approach: Static Routes - topology is linear and nodes are immobile • Problem: Still as brittle as a single string -- can’t back track around breaks • Possible approach: Use AODV, DSR, Roofnet, etc. • Problem: Flapping links will make establishing end-to-end routes nearly impossible • What to do, what to do…? 11/12/04 CENS Seminar 9 Outline • • • • • • 11/12/04 Mexico Seismic Array Deployment: Set Up Interplanetary Networking: Set Up & Relevance Delay Tolerant Networking (DTN): Architecture DTN vs. Email Mexico Deployment: DTN Approach Modular DTN Architecture for Sensor Networks CENS Seminar 10 Interplanetary Networking: Motes in Space • Huge transmission delays – Several seconds to the Moon – 5 - 20 minutes to Mars – 1 hour to Jupiter – 7 hours to Pluto • Lossy links • Long disconnects – Sometimes predictable 11/12/04 CENS Seminar 11 Mars Communication http://www.astrosurf.com/lombry/qsl-mars-communication3.htm 11/12/04 CENS Seminar 12 Mars Connectivity QuickTime™ and a Cinepak decompressor are needed to see this picture. 11/12/04 CENS Seminar 13 Oops! 11/12/04 CENS Seminar 14 Properties of Interplanetary Networking •Links are lossy and high delay •End to end paths unreliable •Interactive / chatty protocols break •Need to communicate across varying network technologies, including non-IP networks –Very high round trip times, –Even for single hop •Links are often disrupted –Node mobility –Node powered down •Asymmetric links •Asymmetric node capabilities 11/12/04 CENS Seminar 15 Delay Tolerant Network Architecture • Messaging service (Like Email) – Insulates applications from network behavior – Asynchronous, deferred transmission – Non-interactive end-to-end You’v eGot DTN! 11/12/04 CENS Seminar 16 Delay Tolerant Networking • Bundles are Basic Unit of Transport – Written as files onto persistent store – Resilient across server / node restarts • Custody Transfer of Bundles – Represents a QOS agreement: custodian tries real hard to transfer custody and not delete until this has happened – Conceptually moves the “end” of an end to end transaction – Custody is asserted, rather than given. • Overlay routing of Bundles among Custodians – Custodians may be separated by intermediate nodes. 11/12/04 CENS Seminar 17 Delay Tolerant, Not Run-Over-by-a-Bus Tolerant VS • Bad things can still happen with Custody Transfer • Brings up the question of trust. When should a node trust another more than it itself? • Risk vs. Resource trade-off 11/12/04 CENS Seminar 18 Bundle Agents and Clients • Client nodes register with Bundle Agents to send and receive Bundles on their behalf – Email analogy: Pop or IMAP client connecting to a server – Based on the destination name, delivery is made to the corresponding Bundle Agent. – At this point it’s considered to be done. Again, similar to Email. • Naming scheme allows for late binding and separation of names from nodes – Determination of which node(s) receive the message may be deferred • Connects different domains via gateways – Late binding names allow different domains to communicate (through gateways). – Connects loosely coupled “Internets” – Official DTN specification uses Tuples: • (Interdomain Label, Opaque Intradomain Label) 11/12/04 CENS Seminar 19 Outline • • • • • • 11/12/04 Mexico Seismic Array Deployment: Set Up Interplanetary Networking: Set Up & Relevance Delay Tolerant Networking (DTN): Architecture DTN vs. Email Mexico Deployment: DTN Approach Modular DTN Architecture for Sensor Networks CENS Seminar 20 DTN vs. Email DTN? Email? What’s the difference? Professor Culler 11/12/04 CENS Seminar 21 DTN vs. Email Biggist Difference: DTN Overlay DTN is able to make progress towards the destination, even when no contemporaneous route exists Cool Professor Culler 11/12/04 CENS Seminar 22 Outline • • • • • • 11/12/04 Mexico Seismic Array Deployment: Set Up Interplanetary Networking: Set Up & Relevance Delay Tolerant Networking (DTN): Architecture DTN vs. Email Mexico Deployment: DTN Approach Modular DTN Architecture for Sensor Networks CENS Seminar 23 Mexico Deployment: Routing • DTN has the “Tuple” address: (Interdomain, Intradomain) – Doesn’t really apply to Mexico – Single domain • Hybrid Routing Approach – You know where you want to get to, but not how – Overlay static on top of mesh routing • Using Roofnet – link-quality aware – minmizes number of transmissions – Use mesh routing to tell you the next hop towards the furthest reachable downstream candidate 11/12/04 CENS Seminar 24 Hybrid Examples • • • • 11/12/04 CENS Seminar Normal case Simple break Partition Reconnect 25 Hybrid Examples • • • • 11/12/04 CENS Seminar Normal case Simple break Partition Reconnect 26 Hybrid Examples • • • • 11/12/04 CENS Seminar Normal case Simple break Partition Reconnect 27 Hybrid Examples • • • • 11/12/04 CENS Seminar Normal case Simple break Partition Reconnect 28 Hybrid Examples • • • • 11/12/04 CENS Seminar Normal case Simple break Partition Reconnect 29 Mexico Deployment: Handling Data • DTN suggests exchanging Bundles (Files) • Divide data into one-hour segments • Augment with meta-data – To, From, Data, Size, etc. • • • • 11/12/04 1.2 - 1.8 MB compressed Store as files on disk (even on intermediate hops) Easy to recover from server and node restarts Human manageable CENS Seminar 30 Node Software Architecture Upstream Nodes Duiker CENS NODE Inbox Drafts Bundle Bundle Bundle Data File Data File Data File Bundle Receiver Bundle Sender Downstream Node 11/12/04 Bundle Forwarder Outbox Bundle Bundle Bundle CENS Seminar Seismic Data (Q330) 31 Routing Data Managing Space • Nodes are equipped with 2 - 4 GB of storage: 40 - 80 days of data • A Bundle is deleted when ACK is received from sink for locally generated data – ACK is just another bundle – Application-level ACK • When space gets low – Node refuses to accept transient data (“route around me”). Better than accept and drop. – Delete transient data when space gets low • No intermediate custody transfer • This policy delays the deletion of original data the longest 11/12/04 CENS Seminar 32 Transmission Priority • Transmission priority affects performance – Want to avoid starvation • Priority can be based on: – – – – – 11/12/04 Age Transient vs. Local Number of hops traversed Number of times sent Some other measure of how hard it was to get this far CENS Seminar 33 What about Off the Shelf DTN? • • • • There is an official Bundle specification There are 1 1/2 reference implementations of the Bundle specification Why not use it? Because… – Only reason why Mars worked, and only way Mexico will work, is because scenario specific information was used. You MUST do this. – It’s impossible to do a good job in all scenarios – The more generic you try to be, the more it looks like flooding. – The more options and switches you support, the more brittle and complex it becomes • See Sendmail configuration file • 11/12/04 How do we remain flexible and customizable, yet stay SIMPLE? CENS Seminar 34 A Modular DTN Software Architecture • • • • • Like Click Modular software IP router Easy to plug modules together Easy to create new modules Results in an optimized router // Declare three elements… src :: FromDevice(eth0); ctr :: Counter; sink :: Discard; // .. Connect them together src -> ctr; ctr -> sink; 11/12/04 CENS Seminar QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. 35 Click’s Elegance • Click modules consume and produce the same data structure – IP packets • • Same is nice and simple. Easy to compose modules if they input/output the same thing IP is nice too (as opposed to something obscure or too generic) QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. – Strong foundation in specifications – Many compliant implementations as examples – People are already familiar with certain features and behaviors of IP routers 11/12/04 CENS Seminar 36 Click and DTN - High Level Click • • • • 11/12/04 DTN • • IP packets IP specifications Lots of existing IP routers Lots of specific IP features/behavior to choose from CENS Seminar Bundles Bundle specifications – • • Very few Bundle servers Lots of Bundle features and behavior to choose from – 37 Experimental and evolving In this area, Bundles have a richer set of potential behavior than IP Introducing this Approach to Sensor Networks • EmStar – – – – – – Component based framework for sensor network applications (on Linux) Lots of services and applications already existing Simulation / Emulation / Deployed Modes Heterogeneous applications (Stargates / Motes) Debugging, monitoring, and visualization Growing user-base /dev/servicename /dev/fusd Transceiver Mote User Transceiver Mote Server Transceiver Mote Client Transceiver Mote • UCLA, USC, MIT, Umass Amherst, Ohio State, etc. • Intel Research, Xerox PARC, Microsoft Research Emulation Array HostMote Protocol Kernel MN MN MN Node Node Node 001 002 003 kfusd.o MN Node N … Emulation Mode Robust multi-process, microkernel architecture 11/12/04 CENS Seminar Visualization Tools Simulation Framework with real RF channels 38 Relationships Between EmStar Modules Unclear Example EmRun Configuration File include link/link.run process mdiff_test { waitfor = mdiff; type = once; noclean; cmd = "devel/microdiff/mdiff_test"; } &link_udp(udp0); &link_linkstats(udp0,ls0,show="leds:core"); &link_neighbors(ls0,show="hide"); process mdiff { type = once; noclean; cmd = "devel/microdiff/mdiff --uses ls0"; waitfor = ls0; } process mdiff_filter { waitfor = mdiff; type = once; noclean; cmd = "devel/microdiff/mdiff_filter"; } How to bring Click’s readability and ease-of-use to EmStar? 11/12/04 CENS Seminar 39 Bolt: Architecture Description Language for EmStar • Makes relationships between modules of an EmStar application explicit • Began as a class project with Eddie Kohler and Todd Millstein last spring • Implemented in CIL and Ocaml 11/12/04 CENS Seminar 40 Bolt Config.: White Board to Emacs Buffer udp0 = udp() ; ls0 = linkstatsd(); nd = neighbord(); mdiff = Mdiff(); Instantiating Connecting component devices Data Flow components mdiff_simple_app [packet_dev] <-> [app_packet_opts] mdiff; mdiff_simple_filter [packet_dev] <-> [filter_packet_opts] mdiff; mdiff [lu_opts] <----------------------> [lp_opts] ls0 [lu_opts] <-> [opts] udp0; mdiff [sc_opts] <- [s_opts] nd [ls_opts] <-- [s_opts] ls0; nd [lu_opts] <-> [lp_opts] ls0; 11/12/04 CENS Seminar 41 BUT WAIT, THERE’S MORE!… 11/12/04 CENS Seminar 42 Click Modules vs. EmStar Modules • • • • Click modules are C++ classes EmStar modules are separate processes Click modules communicate via function calls EmStar modules communicate by passing bits or text via device files • Since Click modules are compiled, they benefit from basic compiler checks and optimizations. Click itself does higher level checks and optimizations • Wouldn’t it be nice if EmStar modules enjoyed similar benefits? 11/12/04 CENS Seminar 43 Bolt: Compiler-like Analysis and Optimization for EmStar • • • It Slices! It Dices! It does static analysis and optimizations not otherwise possible across EmStar modules No user changes or annotation required of EmStar code – Though it could help • Bolt statically infers type-safety violations between two functions in different processes (EmStar modules) – Despite type-obscuring function calls and heavy use of function pointers • • 11/12/04 Bolt uses subgraph isomorphism to identify and swap out combinations of modules with more efficient combinations that perform the same function Analysis not completely sound nor complete, but hard given the circumstances CENS Seminar 44 Implementing DTN Using Bolt • • DTN looks like any other set of EmStar modules What exactly are the EmStar modules that make up the DTN suite? – Click has benefited from the existence of numerous IP router implementations – DTN must build up experience from real deployments (Mexico and others) – Identify and extract reusable DTN related components, for example: • • • • Managing a large number of concurrent file transfers between a pair of nodes Despite large delay and disconnections Recover across restarts LTP (S. Burleigh) & File Mover (A. Parker) – Goal is not necessarily to be able to build a complete DTN solution for your particular situation – Rather, to reuse the ones that make sense, and build from there 11/12/04 CENS Seminar 45 Outline • • • • • • 11/12/04 Mexico Seismic Array Deployment: Set Up Interplanetary Networking: Set Up & Relevance Delay Tolerant Networking (DTN): Architecture DTN vs. Email Mexico Deployment: DTN Approach Modular DTN Architecture for Sensor Networks CENS Seminar 46 References • • • • • 11/12/04 DTN: http://www.dtnrg.org IPN: http://www.ipnsig.org Click: http://www.pdos.lcs.mit.edu/click EmStar: http://cvs.cens.ucla.edu/emstar Bolt: http://lecs.cs.ucla.edu/~adparker/Bolt CENS Seminar 47