Joint Workshop for a Common Signal Processing Operating Environment June 11 - 12, 1996 102, MiRC, Georgia Institute of Technology, Atlanta, Georgia Acronyms ACL: Application Configuration Language (by Mercury) API: Application Program Interface COE: Common Operating Environment CORBA: Common Object Request Broker Architecture COTS: Commercial Off-The-Shelf Technology DoD: Department of Defense GEDAE: MPI: Message Passing Interface OS: Operating System SPW: Signal Processing Worksystem (by Cadence) Tuesday, June 11 0800 Introduction to the Workshop: David Toms, Staff to DASN (C4I) Director, Naval Signal Processing Systems Two-way secure DoD shopping over Internet for hardware Domain-specific software reuse library 0815 Workshop Organization and Process Dave Benett, moderator/facilitator, Dynamic Systems. 0845 The Global Command & Control System COE: Frank Perry, DISA DISA is Defense Information Systems Agency: 2/3 cost of information systems is reinventing the wheel--- GUIs, communication protocols, Instead, build framework to plug in functionality in order to collapse multiple special-purpose terminals into one multi-use terminal COE for all DoD branches Four levels 1. Applications: tactical 2. Common Support Applications: office automation, command center 3. Infrastructure Services: FTP, HTTP, IRC (Chat), News 4. Kernel: Printing, OS+Extensions, Windows, Icons/Menus APIs that are backward capability Build of COE grabs kernel and pieces at other layers as needed Unix: Common Desktop Environment PC: Windows/NT Note that many kernel functions are not needed for embedded real-time applications 0945 Armed Services Programs & Viewpoint (includes 15 minute break) A. Navy Programs -- Tim Singleton, PMO428 UYS-1 and UYS-2 Standards-- 15 years behind COTS Other Navy programs using COTS No common Hardware/Software framework PM0428: Hardware Tactical Advanced Signal Processing (TASP) Need for Software Reuse Library + Toolset and NOT a new standard B. Air Force Programs -- John Hines, WL/AA & Ralph Kohler, RL/OCSS Buy entire systems (airplane) with subsystems Image understanding and Automatic Target Recognition has standardized on Khoros Khoros now has code generation plus control and dataflow models of computation, CORBA-compliant, and hooks to Mercury systems, but no Ada/VHDL code generation Overwhelming validation costs: Matlab -> C -> Real-Time by parallelizing imperative programming onto Mercury Hardware C. Army Programs 1. Jim Hilger, NVESD benchmarking in lab and field they get algorithms from other people written in C 2. Philip Acuff, MICOM, Huntsville RTEMS (Real-Time Operating System for Multiprocessor Systems) Posix-compliant threads in both C and ADA source code 3. Jonathan Sims, MICOM, Huntsville Missle system needs low-power high-speed interconnect for the multiple modules to exchange results processing with high data rates (total power for electronics is 150 Watts) D. DARPA Programs -- Jose Munoz, DARPA/ITO GEDAE Heterogeneous Distributed Computing (funding for Khoros/Matlab, MPI, Myrinet) 1. Matlab -> ACL -> Mercury 2. Matlab 3. Khoros 4. Two-level multicomputer (Sanders/Myricom)--- separate processing and communications 5. Sanders Hardware/Software design environment Future 1. Fault-tolerant computing 2. adaptive resource management 3. low power 4. signaling 5. signal/vector processing library 6. NOW 7. visualization 8. intelligent DRAM 1200 Lunch 1300 Industry Viewpoint 1: System Developers A. RASSP Model Year Architecture/Tools/UYS2A Experience: Bill Kline and Bernie Schaming, Lockheed-Martin ATL 1. Integrated 11 tools 2. Sponsored JRS for architecture definition and visualization tool 3. Sponsor UC Berkeley for mixing control/dataflow models of comp. B. TI's Rapid Prototyping Environment: Les Priebe, Texas Instruments, MSG 1400 Organization of Working Groups Potential topics: COE architecture; application development tools; APIs; library standards (content, interface); etc. 1420 Break 1440 Breakout sessions 1. Need--- identify customers, value-added 2. Requirements--- five most critical, essential vs. desirable 3. Features 1600 Report of working groups Government has 70% market share for signal/image processing COTS 1730 Government-only session 1830 Adjourn Wednesday, June 12 0800 Opening Remarks: David Toms, Staff to DASN (C4I) 0820 Bruce Lewis, MICOM Object-Oriented Missle Software Architecture using Honeywell tools DoME/ControlH control modeling, generate MetaH and Ada DoME/MetaH architecture description, outputs Ada, etc. Ada -> Tartan Compiler -> i960 multiprocessors Hardware/Software Codesign 0850 Industry Viewpoint 2: COTS Manufacturers (20 minutes each) A. DSR's Middleware API: Richard Carroll, DSR Processing systems for towed sonar arrays on-board submarines Using open standards and retargettable software Unmodified COTS Four VME chasis B. Mercury's ACL: Arlan Pool, Mercury Computers Application Configuration Language Built a common layer on top of various operating systems based on MPI, C/C++, event-driven, VxWorks C. Sky Computers Development Trends: Joe Germann & Anna Rounbehler, Sky Similar to Mercury architecture, but Sky goes into higher levels and has better C compilers 1000 Industry Viewpoint 3: Software & Tools (20 minutes each) A. COSSAP Development Trends : Ravi Subramanian, Synopsys/COSSAP Communications Simulation and Analysis Package Software complexity growing at 10x every 6 years in systems Hardware complexity growing at 8x every 6 years in systems Software specification: scheduling, macro pipelining, code in-lining, buffer memory Multiprocessor dataflow CodeGen advisor: firmwave profiling Two-way interface with Momentum Filter Design tools B. RIPPEN : Joe Presley, Orincon C code generation: control and dataflow models C. DSPView: Richard Blanchard, Lucent Technologies PGM simulation and code generation environment D. Autocoding: Chris Robbins, MCCI PGM simulation and code generation environment Distribute Graph Translation Tool and AutoCoding Tool by FTP 1200 Lunch 1300 University Viewpoint: Software & Tools (20 minutes each) Ptolemy Development Trends: Brian Evans, UC Berkeley 1300 Breakout sessions 1500 Report of working groups: Future actions 1600 Government-only session; others adjourn 1700 Adjourn Motivation 1. Signal processing systems consist of DSPs, high-performance computing, real-time subsystems, embedded subsystems, workstations (Heterogeneity) 2. US Government has 70% market share for signal/image processing 3. DoD has no common hardware/software framework 4. Navy fields systems that are 15 years out-of-date Goals 1. Domain-specific software reuse library 2. Middleware with APIs 3. Simulation 4. Description of a COE 5. Integration with other applications Problems with Standards 1. Products may evolve long before standards 2. Standards require evaluation of compliance with standards Highlights 1. Global Command and Control System Common Operating Environment (COE) for information systems on multiple platforms and operating systems; Unix/CDE and PC/Windows NT; workstation-based. Not really the model we want in signal processing systems (e.g., real-time systems don't need window-based functions as a workstation would). 2. Real-time Executive for Military Systems (RTEMS) a. freely distributable under GNU CopyLeft copyright b. real-time operating system with POSIX-compliant threads c. release includes functionality equivalent C and Ada versions d. ported to 68000, i386, i960, HP-PA, and Solaris architectures e. Web address http://www.rtems.army.mil/rtems.html. 3. Mutability of dataflow graphs-- to delete a block, make sure that data it is buffering is processed 4. Change in DoD programs: now they ask what hardware vendors can deliver that interface to a given set of software environments 5. Change in DoD funding: performance used to be the sole driver for funding; now, it's cost vs. performance. Cost is development + maintenance (6 years for the Air Force and 20 years for the Navy) 6. Characteristic curves describing efficiency vs. bandwidth in parallel implementations look like transistor i-v curves regardless of the applications. Bandwidth includes all overhead. 7. PGM is one way to leverage commonality among Software/Hardware/Ifc models for signal processing because one can specify all of these in PGM (ARPA funded SPW -> PGM conversion) 8. Khoros 2 has dataflow and control models of computation and can generate code. It is the platform of choice for automatic target recognition research and development in the military. 9. Matlab is widely used algorithm development: Matlab/SPW interface (Martin) 0. In systems, software size (lines of code) is growing at 10x every 6 years; 8x for hardware (number of transistors).