Universal Inbox: Personal Mobility and Service Mobility in an Integrated Network: Universal Inbox: Personal Mobility and Service Mobility in an Integrated Network Bhaskaran Raman ICEBERG, EECS, U.C.Berkeley


Design Decisions / Lessons Learned: Design Decisions / Lessons Learned Monday 21 August 2000 10:15 - 10:35 Top-level design decisions Rationale for IP-based approach Why an infrastructure based approach? Leveraging cluster-computing environments: Ninja vSpace 10:35 - 12:00 Design of the ICEBERG components and capabilities Signaling protocol for flexible multi-party communication Service creation model Clearing House for resource reservation 13:00 - 15:00 Design of the ICEBERG components and capabilities Automatic Path Creation service Naming service Preference Registry and Preference Manager Personal Activity Coordinator Universal Inbox for personal mobility and service mobility


Outline: Outline Aug 21, 2000, Monday “Universal Inbox” set of capabilities Goals and how they are achieved in ICEBERG Extensibility and Scalability properties illustrated Aug 22, 2000, Tuesday Details of redirection mechanism Example of extension to the Universal Inbox Access to the Jukebox Service APIs for extension Programmer’s perspective Service provider’s perspective Details of scaling measurements


Motivating Scenario: Motivating Scenario


Problem Statement: Problem Statement Requirement Service integration and personalization Goals Any-to-any capability Extensibility: ease of adding new end-points Scalability: global scale operation Personal mobility and Service mobility Communication devices Communication services


Universal Inbox: What is it?: Universal Inbox: What is it? Policy-based Location-based Activity-based Speech-to-Text Speech-to-Voice Attached-Email Call-to-Pager/Email Notification Email-to-Speech All compositions of the above! Universal Inbox Universal Inbox: Metaphor for personalized, integrated communication One of the first applications to drive the design of ICEBERG User sees unified, conceptual inbox of incoming communication


Design Principles: Design Principles Separation of functionality Separation  independent and reusable components Reuse  easy extensibility Shared network services  Economy of scale Network and device independence Needed for extensibility to new devices Push control towards callee In current communication networks, caller has control Callee needs to have control for flexible handling of incoming communication


Use of ICEBERG Components: Use of ICEBERG Components Infrastructure components for network integration Components in the Internet: open model, can leverage proxy and cluster architectures (Ninja) Identification and separation of components Name Mapping Service (NMS) Automatic Path Creation service (APC) Preference Registry (PR) ICEBERG Access Points (IAP) to proxy for communication end-points Advantages of core infrastructure components Reusable pieces; Extensibility is easier: just add to the piece that requires extension


Use of ICEBERG Components (Continued): Use of ICEBERG Components (Continued) Automatic Path Creation (APC) service Generic any-to-any data transformation Provides data-type independence Preference registry Mechanism for ubiquitous redirection Achieves the “control to callee” design principle Naming service Mapping between different device identities Provides device name independence ICEBERG Access Points (IAPs) Provide network independence


Slide10: Bhaskar’s Cell-Phone Barbara’s Desktop Illustrating Extensibility


Slide11: 800-MEDIA-MGR UID: mediamgr@cs.berkeley.edu mediamgr: Cluster locn. Bhaskar’s Cell-Phone Barbara’s Desktop Naming Service Preference Registry Automatic Path Creation Service


Slide12: 510-642-8248 UID: hohltb@cs.berkeley.edu hohltb: Prefers Desktop Bhaskar’s Cell-Phone Barbara’s Desktop Naming Service Preference Registry Automatic Path Creation Service 1


Extensibility: Extensibility Name-space Hierarchical New name-spaces added by creating a new sub-tree at root Automatic Path Creation service Operators can be plugged in Old operators are reusable Set of IAPs New IAPs can be added independent of existing ones All old IAPs are reachable from the new one IAP IAP IAP IAP


Leveraging Extensibility in the APC: Leveraging Extensibility in the APC This extensibility translates to ease of end-point addition in the Universal Inbox


Implementation Experience: Implementation Experience Extensibility Universal Inbox set of features extended to lot of device and service end-points Scalability Components tested for latency and scaling bottlenecks


Extensions to the Universal Inbox: Extensions to the Universal Inbox Step-wise addition of eight different devices and services to the system Each step involves addition of an IAP – for the device/network or the service Each step integrates the device/service with ALL existing ones


Extensions to the Universal Inbox: Extensions to the Universal Inbox


Extensions to the Universal Inbox: Extensions to the Universal Inbox


Implementation Experience with Extension: Implementation Experience with Extension Examples of extension: IAP for Ninja Jukebox Allow service access from voice-enabled end-points ~ 700 lines of Java Also required addition of operators to APC service: MPEG-3 to PCM IAP for MediaManager Allow access to the MediaManager service Similar code-size and effort No other component had to be touched Operators for G.723 Getting codec to work required effort But, adding to APC was ~ two hours of work ( simple API for adding operators)


Lessons learned: What was easy?: Lessons learned: What was easy? Extension to include a new communication service or device Build an IAP Add appropriate operators Effort involved in building a service is independent of the number of networks it is made available on


Scalability Analysis: Scalability Analysis Shared infrastructure components  scaling and provisioning concerns Would like to Answer provisioning questions Identify scaling bottlenecks Three shared core components are: APC Preference Registry Naming service


Scalability Analysis: APC: Scalability Analysis: APC One round of performance optimization Originally, operators were unix processes and one would run for each path Now, operators can be shared across paths Performance for the following operators Null (copies input to output), Toast (PCM to GSM), Untoast (GSM to PCM), and G.723 encoder, decoder Path creation latency and throughput measured as a function of increasing load 500MHz Pentium-III 2-way multiprocessor running Linux-2.2 with IBM’s JDK 1.1.8


Path Creation:Latency vs. Load: Path Creation: Latency vs. Load Toast Operator Untoast Operator About 50ms latency for path creation


Path Creation:Throughput vs. Load: Path Creation: Throughput vs. Load Toast Operator Untoast Operator About 7.2 path creations/sec at a load of 64 simultaneous paths


Calculation of Scaling: Calculation of Scaling On average 2.8 calls/hour/user Average duration of calls (path) is 2.6 minutes Using these 571 users can be supported by a two-node APC service Encouraging since the telephone network uses expensive hardware equipment for these transformations (TRAU at the Inter-Working Function) About 1/4th of this for G.723 decoder G.723 encoder: heavy-weight


Scalability Analysis: Preference Registry: Scalability Analysis: Preference Registry Uses cluster-based distributed storage for storing preference profiles Throughput: 55.3 requests/sec (for dummy user preference profiles) This means about 71,100 users for a single preference registry Clearly not a scaling bottleneck


Additions to call-setup latency: Additions to call-setup latency Universal Inbox’s redirection mechanisms cause extra delay Preference registry lookup About 35ms Path creation at APC About 50ms Such latencies may be high for regular call setup Need to be brought down in the next round of implementation


Related Work: State-of-the-Art: Related Work: State-of-the-Art Commercial services Concentrate on functionality No any-to-any capability Research projects Mobile People Architecture: Personal Proxies Telephony Over Packet networkS UMTS Issues not addressed: Infrastructure support for network integration Extensibility Scalability Personal mobility + Service mobility


Further Plans: Further Plans Extend to PCM end-points PSTN phones, via H.323 gateway Build IAP for interfacing Hypothesis: all existing end-points and services should interoperate without modification (e.g., Jukebox, MediaManager, Two way call with Cell-Phone, VoIP, etc.) Inside department deployment Make system more usable Extend to more services and end-points Scaling and latency issues Services on vSpace Leverage cluster-computing features


Summary: Summary Universal Inbox: metaphor for any-to-any communication and service access Personal mobility redirection by preference registry Service mobility result of the any-to-any capability Architecture viable for global operation IAPs can be developed and deployed by independent service providers Extensibility Made easy by the separation and reuse of functionality Open Questions Security issues Billing issues