1. Clusters Massive Cluster Gigabit Ethernet System Architecture for Extreme Devices David Culler http://www.cs.berkeley.edu/~culler U.C. Berkeley DARPA Meeting 9/21/1999

2. 9/21/99Endeavour Sys. Arch2 Recap: Convergence at the Extremes Arbitrarily Powerful Services on “Small” Devices –massive computing and storage in the infrastructure –active adaptation of form and content “along the way” Extremes more alike than either is to the middle –More specialized in function –Communication centric design »wide range of networking options –Federated System of Many Many Systems –Hands-off operation, mgmt, development –Scalability, High Reliability, Availability –Power and space limited => simplicity Each extends the other

3. 9/21/99Endeavour Sys. Arch3 State-of-the-Art: Very Large Systems Scalable Clusters Established –high-speed user-level networking + single system image –naming, authentication, resources, remote exec., storage, policy Meta-system glue over full OS and Institutional structure –Glunix (UCB), Globus(ANL), Legion (UVA), IPG (NASA), Harness, NetSolve, Snipe (UTK),...GlunixGlobusLegionIPGHarness, NetSolve, Snipe –uniform, multiprotocol communication & access mechanism –personal virtual machine spanning potentially diverse resources »constructed and managed “by hand” Key challenges –Automatic Composition, Management, and Availability –Scalability to global scale –Ease of development for global-scale services

4. 9/21/99Endeavour Sys. Arch4 State-of-the-Art: in the small... Unix-like support in a small form factor + real time seasoning –microkernels dominate »Commercial: PSION, GeoWorks, WinCE, Inferno, QNX, VxWorks, javaos, chorusOS,PSIONGeoWorksWinCEInfernoQNX VxWorksjavaoschorusOS »academic: Exokernel, OSKit, ucLinux, ELKS, ExokernelOSKitucLinuxELKS –+ PalmOS, BeOS,+ PalmOSBeOS –Components and mobile objects: jini, corba, dcom,... => tracks the 80386 –when it becomes ~ 1990 PC Unix will run on it –ability to remove components (modularity) + fault boundaries more important than performance –legacy applications less dominant add-hoc networking for connectivity

5. 9/21/99Endeavour Sys. Arch5 Design Issues for “Small Device OS” Current: Managing address spaces,Thread scheduling, IP stack, Windowing System, Device drivers, File system, Applications Programming Interface, Power management Challenge: How can operating systems for tiny devices be made radically simpler, manageable, and automatically composable?

6. 9/21/99Endeavour Sys. Arch6 Emerging Devices RF COTS Mote –Atmel Microprocessor –RF Monolithics transceiver »916MHz, ~20m, 4800 bps –1 week fully active, 2 yr @1% N S EW 2 Axis Mag. Sensor 2 Axis Accelerometer Light Intensity Sensor Humidity Sensor Pressure Sensor Temperature Sensor Laser mote 650nm laser pointer 2 day life full duty CCR mote 4 corner cubes 40% hemisphere

7. 9/21/99Endeavour Sys. Arch7 Micro Mote - First Attempt

8. 9/21/99Endeavour Sys. Arch8 Structured Communication-Centric System Architecture Active Proxies –connected to the infrastructure –soft-state, bootstrap protocol –transcoding Ubiquitous Devices –billions –net + sensors / actuators –net + UI => flow devices Service Path Scalable Info. Utility Base –highly available –persistent state (safe) –databases, agents –service programming environment Service Paths –aggregate flows (rivers) –transcoding operators

9. 9/21/99Endeavour Sys. Arch9 The Large: Info Utility Platform Not just storage and processing, but distributed innovation of scalable, available services Base Pgm extends the Ninja service platform Complex node ULN service exec. env. Scalable, Available Service Platform Registry PDDS (NB) Discovery automated smart-client fail-over & LB xcode & soft-state via Active Proxies traditional OS functions as services - platform built by push services also Utility requirement => Path connects device to clustered service through Soft-State APs, graceful failover within service via non- blocking PDDS and RMI*

10. 9/21/99Endeavour Sys. Arch10 Key Utility Requirements Utility Service Spreads itself over multiple Infra. Service Providers –persistent state becomes decoupled from service (Oceanic) »preserve security model –contractual relationship between service and platform –SDS, QoS, LB => negotiation, monitoring, adaptation –effective incentive-compatible economic mechanisms Sevices composed from utility serv. of other providers –negotiation arch. generalizes path formation –fail-over across competing services, not homogeneous operations »self-checking, transactional service API –economic mechanisms permeate services Massive information flows –via huge data stores and via vast sensor nets (Rivers) –service-wide auto-scheduling of flows

11. 9/21/99Endeavour Sys. Arch11 The Small: radically simple OS for management and composition Communication is fundamental –treated as part of the hardware, not “the system” Push path concept clear into the device –device fundamentally depends on infrastructure –devices typically have well-connected proxies Focus on scheduling discrete chunks of data movement not general thread scheduling and unlimited memory management –there may be a bounded amount of work per chunk to xform or check data –easy to get very predictable scheduling device network S A S A device network UI

12. 9/21/99Endeavour Sys. Arch12 Precursors to the next generation Operating systems that are not called “operating systems” eg: modern disk controller –event scheduler handling stream of commands from network link, controlling complex array of sensors and actuators, performing sophisticated calculations to determine what and when (scheduling and caching) as well as transforming data on the fly –automatic connection, enumeration, configuration –but several simplifying assumptions must be removed Complex array of Sensors and actuators Network link: - EIDE, SCSI - FCAL, SSA - USB, 1394 - ???

13. 9/21/99Endeavour Sys. Arch13 OS as little more than FSM Commands are an event stream merged with sensor/actuator (or UI) events Discrete flows to/from network General thread must be compiled to sequence of bounded atomic transactions –spaghetti part of an application is configuring the flows –steady-state is straight-forward event processing + signaling unusual events continuous self-checking and telemetry –rely on the infrastructure for hard mgmt stuff push very simple flow apps into devices correct-by-construction techniques for cooperating FSMs as basis for automated configuration and mgmt

14. 9/21/99Endeavour Sys. Arch14 UCB Testbed 1x300 proc + 10x20 proc SAN clusters across depts. integrated through multiple gigabit ethernet extended out throug 100s desktops, RF laptop, IRDA PDA, Cell Phones, Pagers, and numerous motes Cell Phones PDAs Future Devices Wireless Desktop PCs Servers Clusters Massive Cluster Gigabit Ethernet

15. 9/21/99Endeavour Sys. Arch15 Plan Year 1 (Base): –Large: Deploy Ninja Service Platform on Cluster-of-Clusters –Small: Prototype over PalmOS + wince + uc-Linux Year 2 (Options 1 & 4) –Automated service composition architecture –FSM-OS and negotiation/mgmt architecture –Broad simulation environment Year 3 –Deploy widespread services, devices and feeds –Evaluate against high-speed decision making applications