security challenges associated with endeavor project mobile code: mainly focuses on how to protect the remote host running it want to point out that there is an equal problem, how do we protect the mobile code running on remote machine from a hostile host. may require security systems in infrastrucutre, and hardware don't trust applications to be honest, IE bank ledger balance AUotmatic generation of the best security protocols -old system of proof to verify correctness -new approach is model checking, seems to be more of a magic bullet -plug them into a framework, analyze with program, and get list of shortcomings (stevross:this sounds really cool) -once have language down, its easy to do checking. combinatorial generate all protocols of a given size to find the best security protocol with those properties -find a faster protocol -tremendous possibiliteis here -Access Control -still substantial questions raised by access control -based on temporal, or adhoc properties, -completely bogus setup we have here, should have just been able to dynamically gain access to display through wireless mechanism put up display what would keep some student playing quake from taking over display? how do we prevent that rental cars, need to temporarily associate some properties you hav with car, don't want car stolen, need to do with some assurance need to rethink access control with time based, geographical, other ad hoce properties, need a seamless way to do this some reliance on tamper resistant hardware java cards, will be widespread java rings etc, will be increasingly prevalent, but lots of problems come up because of this new ways to attack them, extract secrets differential power analysis, used to read of keys being processed by these devices how do you fit this into very small light portable want a mini-tempest version, provide protection against other types of physical attacks against these systems fips 140-2 standard Microauctions/economics for resource control, very small scale auctions how do we tag information so that rights management information can be restored how do we make applications survivable to hostile attacks agains infrastrucuter want to do all of this and maintain privacy PAT and privacy issues area of overlapping interests alot of what we need to cope with is that this are dynamic not static, its not private vs non-private, its this is my house, normally I want it private, but broken now so I want electronic plumber to reach in to fix it does that mean theres an element of negotiation as part of that? yes what degree do you expose user? undesirable if user had to approve every RPC/open request on data object provide mechanism to dynamically construct groups on fly change temporarlly, on geographic, and time broader mechanisms, native NT file sharing mechanism just doesn't cut it (need to talk to tygar about the identity store, and our acl mechanism, plus active certs) -- project 2 tacit knowledge mining networks of people who interact with each other people who are influential have ties to each other most influential authors are fairly tightly coupled group (say architecture, networking) much more tightly coupled than people outside compute first eigen vector get very strong weights in central clusters from adjacently graph of connected operation not guide to people being smart, but surprsingly good if instead it was citations to journal architectures you find the same structure, authoritites tend to cite each other more duality between authority of documents and authority of people can get in directo communication by tracking logs of people been done in small groups not 100 % reliable, but better than nothing at all not unlike what you would do when working with people at a distance what journals have they published in what people have they worked with were is their expertise dependent on knowledge base working in need to log things email messages access to databases logging document access document access matrix has similar properties take adjancy matrix of bipartite graph and multiple by transpose gives links, and same centrality structure by computing first eigen vector there are algorithmic issues to be applied here, eignvalue is just on of them other algorithms are interesting to explore eigenvalues attractive because they scale very well context, will come up quite a bit -physical location of where conducting activity -other people in the activity -other pieces of evidence that help to define what activity your in the middle of maintain not just document, but whose made changes, whose viewed, indirect data thats not in the document (meta data) should be able to record fairly robustly from log data whats attractive about structural analysis (some difficulty in privacy issues) don't require looking at content of message, just volume and frequency and infering context from that where do they fit into endeaver? props are just one class of internet appliance require ubiquitous access and QoS complementary to other appliance, phones, and wearables some augmentation of cell phone have voice channel, revival of interest in video, wearable computing props, very interesting physical appliance that lives in the world and moves around tacit knowledge mining try to provide information to users that is useful for them to do collaboration also trying to infer what people are likely to be interested in, attention management, shipping data around aggressively, try to guess what people will be interested in by looking at things done in past very good chance they will visit documents that people in the same project looked at good application for using collobarition techniques to improve prefetching etc activily working on PROPS right now with funding some delivarables for first year want to work on later -indexability -organizations theory -games/exonomic theory -data mining/m.c. clustering -information retrieval - on-line optimizations - search Given a workload, profile of data you have, and queries you expect turns out that every workload has a quality called indexibality neet trade off between redundancy and overhead you get at access time Actually, in some interesting cases, can precisely use this data. point is that we have been strictly interested in performance in the past, now we are hearing that SAM is replacing performance should look into that Organization Theory ------------------ What are principals that make organizations and people behave the way they do Games/Economic Theory --------------------- How to integrate this with insights, example, game theory how a group of players that are a little less smart, suddenly they become happier Data Mining ----------- how to apply economic principals to data mining and clustering look at ocean of data, have specialist try to find something intersting consumer a who buys this stuff also buys other stuff inject economics into system to determine what data is interesting, data that is interesting is data that will affect my performance |Policy compiler: what is istore trying to accomplish what rules, use economic data mining to get this (sorry batter died, note may not be totally acccurate) information retrieval ------------------- mine web, which documents are important determine just by structure on-line optimization -------------------- database, new flow based models, intersting techniques to be used there search -------- occasioanlly to solve combinatorial problem search huge solution space, many people look up algorithm paradigms -genetic algorithms -simulated annealing etc, create a system that understands all these tools and more, and uses user feedback to optimization parameters that make or break this search effort need 5 great applications to test this on -optimizing optical protocol for security - 3 year plan, 689 soda talk to atleast 3 groups over a couple of days develop relationship, here is where economic theory comes in conflicting interests your best hope is I'll come and say oh I know how to do that thats my worst nightmare what I'm really after is to find out by interacting with endavor folks the new algorithmic theory problems, that are more important Star Trek - voice interfaces / better ui - locator people (soda, or cory) just go to web page, says where you are get report card for how available, collobaritive you are each week we interface, task collobartion, bldg access slap it just like star trek ;) Big Picture weather, probably the easiest, have the hardware to do that around the bay easily anything can do with 1000, rather than 12, look into importance of abstraction Mead Conway and Uniform Building Code putting constraints makes it easier to compose and build grander systems What hardware needs to do to get used - exist has to show up - stability - Type -B feedback, changes, design review - get built into grad classes what will be available by end of summer? depends on what interest is? you tell us what you want probably two or 3 sensor networks working with wireless capabilities interms of range etc 50 nodes, 6 months, 100 dollars webcam in every class room rowe bet he couldn't do it in 6 months data streams that you can look at atleast maybe even targets for code need to find way to make reprogrammable over data link Books that everyone in endeavor should read The Diamond Age Stone Crash, stephenson (he has 5 copies in office if want one to borrow) True Names, vinge Made to Measure, philip ball beyond papers that ack endeavor group sabbatical don't teach for a whole semester focus graduate teaching/taking violent change, put revolution back in berkeley endeavor symester in spring 2000 Endeavor course with projects tutorials and stuff over the summer to learn for it david is talking about serious shedding of commitments, should be able to do that in 7 months when people call you to do things in jan next year say no, I'm on an expedition take it seriously, turn other things done for this once in a century opportunity to really invent the future staKe our reputation on building the future reinvent the content of your course for that symster focus on endavor, lectures and projects on that focus intense seminar, faculty study group in spring seems to be key faculty need to be more flexible in crossing projects across classes volunteer 252 to have projects that are device related did it before with 150 and 152, takes about a semester to do the homework to make it productive how do we get to where spending 20 hours a week, not 2 hours a week thinking about these problems how can you revolutionize the world if you only think about it 2 hours a week somehow need to facility this environment in soda, need to use interaction spaces really force our selves to go out there take back space from facist undergrad students why cant we have a prop on network on 4 floors of soda that you can connect to, infrastructure for people to connect to it project make resources to invest to make this infrastrucutre built challenge is "demo or die" if don't have that may never get there software people need clear goal of what they want to do to, ie locate any people in endeavor as long as wearing badges utility side, that will be there, where is set of places to put our effort automatic code generation would be nice milestone like endeavor semester would be a good thing to have to focus on getting ready for it truly ubiquotous wireless infrastructure through out soda 10MB wavelan by feb 2000, promised by culler would really motivate data utitlity stuff negotiating trust with untrusted code problem cannot trust the software most people think of other peoples software, not true, its any software must assume malicous from unkonw even known sources can be malicous people to quick at saying use cryptography not only external properties that we need to worry about must also think about semantic properties, what software actually does orthogonal things, require orthogonal solutions cryptography is very good at handering non semantic properties doesn't invent trust from nothing, good for transitivity of trust don't want to rely on model that few monopolies have authority to sign need semantic tools that worry about what code does security monitors, trusted been used for long time has serious limitations - run time cost is large - opportunity cost - runtikme checking alone not always sufficient might be able to detect after 10 seconds that nuclear reactor is not shut down, just detecting is not good enough to stop execution, cant stop time need cryptogropy for external properties need runtime checking but also need static checking -- should help with performance two technics proof carrying code try to do static checking on piece of code not meant to be understood, just meant to be run ask whoever wrote code to send explanation can be understood and can be checked against code advantage: breaks limits of static checking once have proof, only have to check the proof if can actually do this, even for machine code, don't have to rely upon compiler afterwords, only thing you have to run is proof checker (vs. JIT) How do you produce this proof? -something the code producer should do -most of the time rely on tools to establish properties -java compiler guarantess type safety tool actually knows -other language - gaurantess lack of race conditions -idea is make compilers smarter, so they argue why code has properties expected that way don't have to trust compiler, can run itself current state of art proof is solid piece of architecture proof 10 - 50 % size of the machine code, without compression checking speed is not an issue once put in more complicated properties proof size will grow, although speed should stay the same willing to support as a type B part of endeavor has one compiler for java, that supports dec-alpha, and a subset of C programming language kind of painfully small subset, but is complete puts in enough runtime checks to produce proof working on better compiler for java move to start using off the shelf compilers should be able to do it as post processing stage based on hueristics goal java code to bytecode or machine code, or low level virtual machine on which you can produce optimizations three year plan same thing for C even though C unsafe language, can still extract reason its type safe from C program sometimes fail because programming erros sometiems fail because out of programming style hopefully away to null out pointer, to convince non dangling, or maybe just needs runtime checks someday typechecking will be really be taking for granted "If your program crashed then you gotta change your motherboard, its not the softwares problem anymore, its the hardware" had a hard time coming up with list of milestones did find some critical cross cuts (see slides) lots of different places that have preferences and policies that get specified have some form of specification for policy observe whats happening in system use that to adapt the specified policy very hard to get policy right first time in static specification need common language for specifying (see slides) activity tracking if system tracking everything about you how do you keep them from learning things about you that you don't want them to need privacy policy to hep with this project cross cuts what is a ninja service, what can you expect with it very strong notion of typing, I want something that maps voice to text maybe I want a german language converter maybe I want one that works in 10ms latency stuff not kept in types not dealt with in current version of ninja inquiry, can you do this? no I can't do it but I can do something else, what is language how to do it given policy and specification can you verify it is correct screw simulation (John osterhout) hard enough to build simulator should go for real thing some things easier to do in simulator key management, key distribution, key authorities, who can you trust what is the trust model (tygar) ability to do extremeley complex policies is there (address book in cell phone) but its just to hard to do, theres gotta be a better way big UI ease of use issue self spy tool, what can I learn about myself by using this system, but you could use that to spy on someone else too, small group of people who really understand architecture and how all pieces fit together, might be good to get an overall integrator for whole project, hows it going to happen so divers and majorly complex, function that is needed invite others to retreat? arent we really competing with other groups, wouldn't darpa people be in heavy duty cell mode, probably not a good idea adaptive systems Enabling technologies - pattern rec - graphical models - reinfocment learning - computatoinal economis applications -lightning prediction -machine diagnostics -bioreactors many failure modes, very few observations of failur look at working systems to try and infer failure modes -fiance -bioinformatics -geophyciscs -error-control coding, source coding, -jordan has consulted on applications in all of these areas -State of art in pattern recognition -classification and prediction, take in data stream, face recognition, faces vs non-faces -take large amount of data, and put down to small amount of info you need for decision -support vectrs in data set Osuna 1999 (smu paper) can make accurate detction turns out global optimization problem with single optima, much better than nueral nets, and other older approaches, Technical Challenge : not online training, take a batch of data and massage it is it simply computation? needs 100,000 data points, base dimensionality of 1000 these are called kernel methods, promote method up to space that is flat, project up to higher D space where things flat, being widely applyied these days Layered graphical models, state of art digit recognizer USPS digits built detcotor/ classifier that classifies digits. graphical model, ---- redundancy redundancy pixels --- at 4% with this architecture, best is 1 % wich is basically hman one cool thing about this system, still worked with only half image, and was able to predict the other half of image with high accuracy compressor, classifier, predicter temporal graphical models, -used on wall street to predict markets time series model in which multiple distribute processors cooperating to predict outcome, ie, what is currancy rate, what is approximate sufficient statistic? reduce by orders of magnitude carry along in time, can give performance techneecs applications in speeach, finance, machine diagnosice technical challenge, semantics in endeavor project: - working on learning in vision with malik - working on signal processing iwht EE people - my door is 739 soda and always open - will keep it open, listen and try to consult. intend to offer course every other year and tune to needs of community spirit of this work is never aggregate data, might have good synergy with smart dust oceanic data store motivation none of many devices we have communicate in consistent way widespread mobility of producers and consumers want data in your car area network serverless, homeless world is an ocean of data, anywhere, anyone can get at it truly nomadic move data to where its being used promiscous sharing/caching of data tacit information allows you how to figure out where to move it intellegiently dissemination of info. no longer tied to particular server, multicast if need to really a utility, federation of large, small, all serving data get one uniform view set in coffee shop that has 1000 terabyte server to serve clientsw when sit down, data migrates to local server somehow cafe gets money back for that, economic model built in why a federation? why not a monopoly want to enable everyone, big and small companies, tie together in some way store coppies in some way that has low probability of failure scalability is a huge shallenge how do you make everything act together to give you a coherent view performance is easy to destroy in this system by smallest wrong thing have to be careful to not let that get in the way security no longer a bolt on, its not optional, have to assume it from the very beginning because of data flow confict resolution is what really matters in consitency, acid semantics is just a really strong form of conflict resolution acid, is really more prevention disconnection is resolution All communities only have caching along path from client to server using resource close by but not directly on path is tricky is economic model really good for global optimization? People hoarding data, rather than moving around enabling technologies computational economy - utility provider model, just like electric grid way to get performance is loosely consistent federation of indices check local index, then next one in hierarchy but there is one global source naming & security scheme watch what you do, go to coffee shop every day, mine tacit info use as hints to pre-stage some work in mobile community -- ie flying to washington calendar has alot of useful hints for how to stage working set two issues experience with calendring system is that individuals dont like to put it into online system, social problem has to be overcome is it I like my program or am I afraid to put in system some people do not want others to see there calendar or have opportunity to schedule in their calendar human task support, verses compuer interaction don normans book, the invisible computer, easy to design, only if you ignore fact that people will use several of them do same task on different appliances calendar: phone, pda, desktop want a basic underlying metaphor that might be the same hard to take advantagie of different modalities, uses for devices hands busy cause using tools, eyes busy because driving, speach can't be used because talkin want to be able to use that context info to generate right ui etc maybe everything will converge into palm 23, but still want to use a different interface depending on the context of the use still no real high level tool for speach, none allow to think about targeting application for widely varying platforms, model based designs, problem is designers work better with concrete examples, sketch out scenarious of real design ideas to cmoplete context-aware widgets: now really simple, can we get more, ??? more complex context, what tools am I using with hands, how do you describe those? use tools to extract model, do it for differnt contexts, different devices know context based on calendar, sensing talking in environment, sensing location through GPS? speech corder, futuristic pda with speech recogition, done out of icsi cross pad, captures, context person, time meeting etc need speech recognition expert in your group to help you do this design how do you extract model from rough representation build context aware applications symbolt pdas, have built in wavelan, and bar code scanners to pick up location and other context about objects see what people will build in this new world need way to extract useful stuff from data streams, dust, video, whatever key is to evaluate usage in target settings like classroom and rapid decision making 1) Omnipresent Visual Aid - vision is our dominating sense - two projects which have tyin to endeavor person centric view lets just assume that everyone has a camera that they can carry with them\ what can you do with it? right now, just tourist, shooting pictures, one wants camera++ camera, gps, accelerometer assume computing, communication, etc want it to answer queries for information for you examples have images, additional info (gps), and some kind of context - tourist guide where am I? what is this building? some museums have this, but what about the entire city of paris, want to just point camera at something and get appropriate information using all information helps with context and making problem more tractable - ecological field trip what kind of plant is this, whats this part called, ties in work with DLIB, cal-flora again, same visual image can give you different guesses, because your in California vs. France, vision in context - aide for blind people - usual sensors they have are proximity sensors - vision is distance sensor par excellence, thats why so evolved can get big picture at a distance want to be in situation were solutions that are not 100% are still useful DLIB query example, tiger finder can I eat this plant, will this animal eat me, biological queries ;) are partial solutions with meta-data useful? 2) Second project environment centric -senors are duality, is it carried around by peson, or is it in the environment? smart spaces track and monitor people simpler set up, cars on the freeway -want to know state of traffic, cameras on freeway, collect data 1/2 mile apart find individual cars, and track over time -then reasoning/inference is involved vehicle 2 is changing lanes 50% probability drunk driver from driving pattern can do this real time, and are currently doing field test -why is it easier to track cars? -because they are rigid objects, don't flail around - people much harder because articulated objects track limbs, articulte graphical model much work to be done to do people in real time instead of requiring people to wear tags, or identify where are much simpler ubiquotous technology you could have, infer what the right thing to do is secondly data mining, tacit inference what are activity patterns at different times of day high quality multimedia interaction -multiple programs, all rich data -broadcast of MBONE people know what your doing, public visibility -figured out why hard to do, develop software and change physical infrastrucuter in rooms alll archived on the net, can look at them if could want to have multiple streams when you do this your looking at me talking and the screen want to be able to do both of these at same time +transparencys video tapes multiple sources of information mean it makes sense to use multiple streams capture material, automate sound bytes of material after the fact has to be low cost operation, elliminate labor, build intelligence into room, so instructor can walk in, turn on the lights, and they are on the air trying to automate production of broadcast, by sensing things in room tradeoff, labor intensive vs. low cost how do you make it low cost, but high quality take out media types if trying to produce title, enabling technologies - make the room smart, give you good cues on what you want to do ie, speaker puts in video tape, do the right thing automatically someone in the room asks a question, spatially locate using microphones, automatically position front camera on user, shouldn't you be able to catch everything, and just dislpay it should it be hardware, or what they display when, produce three broadcasts, low bit rate MBONE high bit rate experimental network real networks feed all moderated broadcasts, should put unmoderated streams out there and let people do what they want with it right now keeping multicast working is a major problem in itself, implemented parallel video effects system on the now - tv guys figure out how to create visual intereste -titling, side by side pictures, keep your intention -useing now to put this into stream on the fly -layered coding and transport -simulcast at different bit rates, want layered multicast technology dont mechanise, actually automate it how do you record, archive, and make it useable to the rest of the group, hook in with notepals, could add alot of value more configurable smart space would make things much easier Smart dust so what? figure out whaat to do with everything sensors we currently have light temp pressure humidity magnetic acceleration dont ahve --------- sound images/ video should state of art device, about size of dram next generation package, with biderction rf about the size of a CPU trying to stick everything into cubic mlimiator want to use line of site to meet small size building rf sensor networks, were connectivity more traditional will probably be useful in shorter term can you make microphones? could make hearing aid microphones, compact, limited power, impressive sensing and computation good bar to judge these things by can do high quality digital sound for the kind of power requirements could also get sugar cube sized video camers so think sugar cube sized with rf communication for the kind of things, this is basically off the shelf now, battery is probably the biggest component make a difference technolog is MEMS problem MEMS is making it worse now have all thes devices, so what, how do you get them to talk to each other, really augment reality, and make life better right now, no idea what to do with all the information when it comes back plan: build a ton of hardware right naive software get cs excited about righting software preliminary software already done network simulator give some connectivity, download protocol very quickly verified that DARPA ideas for this were crap and would never work matlab simulation - through out millino sensors, have no gps - don't know where are but do know who can hear them - huge matrix, figure out distribution in 2 & 3d this is just a vision problem, except randomly distributed -using vison algorithms for motion flow and tracking - working with sequin to try and use soda hall model to generate line of site matrix for actuall building if walking around wth pilot, whats your connectivity as you move around the building location could be really neat think about classroom tracking, micropohnes, could be neat 10 m communication IRDA, radio, huge array of gazillion motes that you can access with different modaliteis thousands of sensors in cory and soda people, ID, HCI doors coffee pots, cups cameras volleyball net nefili glove, in 3d acceleration traces, get all that communication sensing and power into grain os sand in tips of fingers tables/conference areas/chairs what tricks could programers use to extract context from this stuff books?missed anthony very much, its probably canny and anthony that will be responsible summary of discussion among graduate students very important to think of apps as coupled with sensors -weather forcasting -app people arent looking at that yet good to have some vertical discussion groups to -couple application and sensing -mobility of data and infereing content -traffic monitoring, good infrastructure no novel application thought of -medical sensors/applications -any volunteers to injest sensors -chronic sensor implants in the human body -virtual keyboard with fingertip sensors some interesting work using input dictionary, and grammer for this as input device -group in japan has done tests with rings that you wear -output devices, what kinds of ubiquotous things put out in world to be put in new context simulating and controlling things has there been enough thought about output device issues?? step outside of model of just information resource, more of a medium to interact with the environment smart rooms has some of this built in Pieces that lined up well Iceberg PAT preference lists Digital Library Project idea of personal annotation non monolithic view of documents, things that can be arbitrarily extended by people probably a good model for data in the future forward references to other papers etc Data Filtering vs. Data Fusion -interest metrics, based on unexpected things happening -better can model data, better can compress it - filtering, data reduction, data fusion data fusion: lots of different data sources, coherent information Two themes that emerged as vertical slices 1) Moving data Prefetching, anticipating where it would be needed and by whom kubi/dave on storage joe/canny on schema, graphical data structures for data representation 2) Context Came up in many flavors several times Localization: localize activity figure out where you are and make decisions leverages off of tracking stuff, maliks camera tracking stuff Milestones Implement the multicamera peolple tracker for -- jitendra Remote discussion section - use props use a robot GSI, do from mobile devices have student names if actually tracking people student names over peoples heads action items: clear calender for 2000 systems seminar 294 faculty meet on friday afternoons have an end of summer meeting (forcing function) problem, configuration, software degradation, management will be a really big problem when more devices dunno solution, but have some ideas how to design components so that they naturally work this way examples: Get out digital camer, says look, I just have short store find component that offers persitent store, negotatie it wether willing to accept it as client persistent store looks for hierarchical etc, how do you sepecify terms, conditions, groups, federations how do you adjust: Real negotations how do you do something like contract/compliance monitoring example: DNS service, the other day machine performing slowly what had happened, dns server got turned off clients tried the first two etc, third one on list worked appears to user as the network got slow could have used adpatation here to use that as primary service renogotiation to find something global in area and adapt taht way how do you make it completely trivial to make clients all work that way with service interaction. want adaptation little motion sensers on fingers, maybe it automatically figures out its the thumb guy, and that watch rely station uses to communicate with someone else figure out who they are automatically lots of diverse stuff now, but none of them really doing the right thing state of art today: huge complicated infrastructure Eric is the only one that used corba lots of work on agent-based negotiation, that don't understand JNIN lacks: figure out who you are as part of federation, no real notion of adaption, what is specificatoin of languages for specifying services, negotation real question is however do it must be really really easiy compliance monitoring will be key, perhaps register with compliance monitoring service boilerplate contracts that people can use as default, most of work just selecting the right contracts and parameterize them te way that you want Related issues: ease of use very important nince to design language that could understand adpative properties potential for lots of interaction between digital library and indever still trying to figure out what those are dlib document client + collobration properties how can make that work well with endeavor rights management: give up content on various terms, lots of negotation, these things might carry over word for word to rights management space really good place to look. wide spectrum of interaction tension between loose contracts and monitoring vs type based, publicshed interface and contracts first use type system to get in door, than have someone to check that semantics match as well question of contract specification, look well? language that says both of these things is possible analog to uniform building code, mechanical, electrical, all interact these can adapt within these gudlines, enough withing regime that adaptation mechanism can work then can prove things about applications. if we don't solve problem one way or another everything else we do is doomed Packing for the expedition lewis and clark, alot of really important stuff was done when they were figuring out what kinds of canoes to buy, etc Millenium, a nother 10 million dollar cluster to build, good platform for endavor -integrate fast, direct data transfer into java/ jaguer work, flexibilt of mobile code, direct acces to network devices postpc environment is moving forward will have it with commercial technology over the next year Problem 1: What is architecture for Billions of Devices to plug into -os for deverse -negotation -vast data storage probably another historyic operating system can build interesting small oses on those systems os should be communication centric, rather than grafted on at then end, of course its connected, that should be started in beginning tremendous confluencers at exterms whats the simplifier, whats going to make the thing your proposing fundamaentally simpler than it is today Software perspective, I want a program that will deploy itself across a spectrum of devices we havent concieved of yet what happens where how does it adapt, cuts accross those options easier to adapt within framwork rather than organize form scratch Simpligying concepts think theres alot to fluid model with little bit of self check 1) Siplifier, theres nothing but a big reservoir all devices dip into and take out pieces composition of cache coherence protocols, very import idea that is comming across sgi integrated with r1000 etc protocol adpatation at every stage we now understand alot about consistency models consistency doesn't mean waiting means promises about what will happen in the future naturally adapt withing hierarchy, adaptation to access commonplace, lmited by diping rate HotSync is crazy, instead think that you are always connected Proxy, talked about alec woos project downside of this view, worlds most unstrucutred model of programming acid work for seperating apart apis, might give room to develop both, take big world into small world operating system here is nothing rather than a protocol engine, raise it up one level becomes totally invisible just want acess to objects FLOWS all about flows, rivers huge disks post pcs smart rocks what is os for smart rock, in that since, just flowing data from one place to another, how do you simplify os if just moving units of flow is that a big simplifier? framework for adaptation, think of resivour, devise pulling out of it bandwidth of pipe is function of how fast you consume seen in river work in the concept of disks everyone pulling does a really nice job of matching flow to capacity of device how do you handle flow into the reservoiur? whats the goal, where are you relative to that calculate error and make simple adaptations wuite easy to formalize a world that looks that way, just flow equations, havent really developed any formalisms in that world might be able to get much more operating systems in that space smart rock/pilots good platform hard part is communication Self-Checking, is completely wrong, dont want self test, want a printer scanner that nows it printed ps correptly, every where in the system where possible it should check heres goal heres action, did it achieve goal if not why not? potential simplifier simulation tool: for many small devices -this is way to get to class B, use similator to leverage it -want distributed simulation for small devices X copilot for 1 million evices -prototype cache data pad everything you can do on pilot with out ever syncing - smart rocks, flow model OS production quality Simple OS small set of building blocks formal system through composition Katz Smart spaces in soda technologies that will make it happen ninja PAT etc Template for Activities mobile code much sager to move xml around then mobile code universal interactor : todd 326 soda discovery process daimler/chrysler: same thing inside vehicles helen: pat preference management etc large scale prototype in industry with some iceberg stuff sprint/daimler 326 Colab, not actively being used, good place 466 playpen idea for smart space deployment questions: patterson: prototypes: build when to test ideas vs build one that others can rely upon to get there work done katz: not everything will be type A, but important to understand smart space infrastructure is depending on building on NINJA so ninja is type B rely on it. Iceberg/Smart space will be type b to people building smart classroom Tacit Info Managment ?? Dunno yet Culler: Deep influence on what we do, all pieces depend on have to work, mine just have to do benchmarks, have to look out for cyclic dependencies its note 15 people seperately, the sum is more than the pieces, must be lots of type B Patterson: hard to know if type B. if you want to be type B you have to do support Katz: also have to be flexible to incorportate good things in industry, industry working in parallel on many pieces as well, must be flexible, adaptive. Telegraph: A universal System for Information Regres, B-1 (read those papers) four layers storage manager: type B will be part of Ninja, transaction if you want, will unify filesystem Database and Web Build on River, adaptive data in clusters Global layer, federate systems that don't work together Query/Browse/Mine data Synergy, Ninja,Gist,ISTORE (see slides) Storage Manager: Historic chance to start over seek vs scan, ever growing memry, alot changed in 20 years units of data variable in size, segments main memory indexing, persistent vs on the fly extra cpus on devices, can you use them to do ndex on the fly, etc revisit and clean up infrastructure for transactions clean api: unify DB/FS/WEB, arenty they all really the same thing? wouldnt it be nice to use same interface to storage River: Share nothing cluster all dataflow operators, no blocking ops, streaming IO model load balance, micro: one disk slower than other macro: plug in new computer it just works flow to where the power is adaptivity more important then perf, need to be able to manage traditional, first you have to lay down the pipes and the dataflow, once setup you can't change it, so you must optimize it idea here is opposite, should eb able to adjust itself from feedback through the dataflow, along path of least resistence one of first steps handling streaming infrastructture. note jsut fast, available consistent, more important cooperative model of adaptivity vs. competitive / economic model which is better for which situation, different parst of system will be different how do you know where data is, and who can compute what, global metadata issue, what services are there Application, how do you find what you really want when you have all the data in the world in front of you ui, ai, logic, merge all of them into queries interaction is a datastream flow, redirect in various ways continously refine, ie multiresolution output system learns from you what you really want with user in the loop different view of data mining SAM scalable, avaialable, maintainable this is rejecting server as building block for storage, reinvinting them connect disks over hierarchy of busses, SAM was largely ignored in the past Faill Fast theory is flawed, nothing really breaks that way, nothing fails fast today, sys admin must observe behaviour and recognize in advance to avoid disaster, humans, go to dentist, take vacation, miss signals even if people looking at it (see web page for details) embed microprocessor with every single disk, use halfhight case, low profile disk, + small pc board inside there intellegent disk brick, Redunant NICS for SAM Diagnostic Processor on the boad to monito the board, that makes it novel device in itself first prototype 100MB eithernet, chasis, with 64 disk bricks that is building block introspective system: hardware enabler to look at system -kubi policy statements, rule based approach as highlevel interface to managing system Make maintainability the #1 goal ISTORE 0 cluster of PC with switches concerned about software availability, adaptability genetic diversity of having bsd type operating systems with same api, see if what happens, genetic diversity learn about adaptability theory from Micahel Jordan and his students, think it will be really different, better new different way to building systems construct SAM benchmarks, got any suggestions, not sure what that means yet. longterm view, use IBM microdrive, about same size as the pentium chip in 5 or 6 years its probably 9 gig, and could put 10,000 in a rack could have a 10th of a pentabyte in a rack ******* Policy Capture, Preferences, Start with Specification it learns what you really meant by observing system and your behavior in reacting with it, this seems to be an important cross project thing ****** rule systems seem to be a scare thing to bet on, dont have a great track record. -some evidence, in maintingin, much simpler world than the AI world. -know what it is to make them safe. hoping that there is some great theory, hoping concentrating on Maintainability give us a much better shot. Automatically configure to all possible performance, harder thing to do, not possible interested in ideas on what will make software better will clearly do hw/sw co design storage manager API from Brewer/Hellersteing, would be good **** should agree on O.S? if tell API then maybe O.S doesnt matter is there a lowest common denomitantor that you can use api design good to design to two o.s. reliablity hybrid is good want op sys to run on PC, gives best software base, hw designers biggest problem is getting applications. **** wants to by truly type B software really tie things together want to survive endeavor focus on getting stuff have to work in way that makes rest of project work overal ninja strcuture and extension of ninja ninja is the way you right the 24x7 part of whatever you do infrastrucutre collects info about disks etc, RPC with money is just using differnt base class corba com problem, no way to gracefully recover from node failures ninja services handle this, service won't fail let you down, atleast one node will be up 100 cds 6x can lay all simultaneously services that map sounds from CD to real name/track of cd implicit is notion of authentication, can control what track you listen to, can enforce whatever policy mechanism to allow pay-per-use policy what keeps people from going outside of model? can buy cd, makers cant keep you from copying it, thats ok, depends, alot of people dont think its OK missing point, I can enforce what ever policy I want within framework, moment you move outside and consider arbitrary actions done to objects cant enforce it cant control what you do with data after I give it to you can control completely how I give it to you home base -only place that persistent state is guaranteed -greatly simplifies everything else -lose device cant loose persistant state simplifier voice enable room control text to speach service every service has device realy trades stock with realmoney, use with pilot and is secure NinjaFax, anyone can call, not anyone will get served driving home point that authentication built in universal inbox, first version due by june 30th by fall versin that other people will use applications were good drivers forcing us to get this stuff right Ecach mint psuedo anonymous cache, Ian Goldberg Endeavor will drive ninja to be the right infrastructure, thats all I expect out of it.