SpaceCubeX: A Path Towards Hybrid On-Board Processing Architectures
Ma hew French and Andrew Schmidt – USC/ ISI Thomas Flatley – GSFC Carlos Villalpando – JPL June 23, 2015 Mo va on: NASA Earth Science Missions
• Instruments produce essen al data to help scien sts answer cri cal 21st century ques ons - Global climate change, air quality, ocean health, ecosystem dynamics, etc… • On-board processing ~100-1,000x than previous missions (compression, storage, downlink) - Current/near-term data at rates >108 to 1011 bits/second • Adding new capabili es such as low-latency data products for extreme event warnings • Missions specifying instruments with significantly increased: - Temporal, spa al, and frequency resolu ons à to global, con nuous observa ons • Mee ng inter-mission reusability goals
ACE GEO-CAPE 3D-Winds Hybrid architecture is a key cross-cu ng technology directly applicable to missions recommended in the Decadal Survey ASCENDS HyspIRI PACE
ICESat II LIST PATH
2 SpaceCubeX Project
• Develop high performance on-board compu ng capability for flight missions: - Leverage heterogeneous processor types (COTS à Rad-Hard) - Incorporate industry experience with commercial device specializa on - Improve computa on/energy efficiency - Support persistent, long-term mission opera ons - Provide poten al infusion into (P)ACE, HyspIRI, GEO-CAPE, ICESat-2… • Address key characteris cs to support mission design: - Size, weight, area, power - Mission capabili es, cost, risk - Risks of supply chain disrup on - Benchmark compila on for robust characteriza on - Commercial tools/compilers interoperability
3 SpaceCube Family
4 SpaceCube Next?
Mul-. • Mul -core processors more easily provide: Core' Memory' IF# IF# – General OS support
IF# IF# IF#
– Sensor' High-level func ons DSP' FPGA' – Coarse grained applica on parallelism
• Co-processors then provide: IF# Mul$% Mul$% IF# IF# IF# IF# IF# Core* Memory* IF# Core*
– Customized accelera on IF# IF# IF#
IF# IF# IF# – High throughput IF# FPGA* IF# Memory* IF# FPGA* IF#
IF# IF# IF# – Fine-grained data parallelism IF# IF# Sensor* Sensor* • Sparked ques ons: – What devices to use? Mul-. – How to connect the devices? Core' IF# IF# Memory'
IF# – How to program the system? IF# IF# IF# IF# – (and many more…) Sensor' IF# FPGA' IF# IF# DSP'
5 Hybrid Architecture Device Candidates
Category Candidates Mul -core Boeing Maestro, Tilera Tile64, Aeroflex LEON4, ... FPGA Xilinx Virtex7, Altera Stra xV, Microsemi, Achronix S22i DSP BAE RADSPEED, ClearSpeed CSX700, TI 320C6701 ... Memory Atmel M65609E, Honeywell HXSR06432, ... Interconnects PCIe (Gen2/3), XAUI, RapidIO, SpaceWire ... Topologies Bus, Mesh, Ring, Crossbar, Bu erfly ... Sensors I/O RS-422, Ethernet, SpaceWire Direct I/O, I2C ...
6 Tradi onal Development Process
Tradi&onal*Hardware*Development:* Benchmarks' Manual'' Topology'' Hardware'' Specifica.ons' Component'' &'Netlist'' Performance'' Implementa.on' Analysis' Crea.on' Results'
SpaceCubeX*Heterogeneous*Hardware*Development:*
While number of processors available to aerospace industry Specifica.ons' is tractable, number of combina ons of these elements and number of interconnect topology permuta ons makes this Applica.ons'' Compila.on'Environment' process difficult and error prone for a human to do manually. &'Benchmarks' Designer' Evolvable'Testbed' Inputs' Flight'Hardware' Architecture'' Performance'' Genera.on' Engineering'Hardware' Results' TRL* Hardware'Emula.on' 7 Design'Explora.on'' Simula.on' Feedback'Loop' SpaceCubeX Approach
• Develop avionics architecture and generator suite: – Candidate device analysis – Evaluate with benchmark applica ons • Profile architectures in simula on environment and select leading candidates for emula on testbed • Implement architecture with selected hardware in emula on testbed • Evaluate implementa on • Assess results and lessons learned
8 SpaceCubeX Development Process
• Updated to support hybrid architecture with focused design space explora on • U lize genera on and analysis tools to eliminate manual fit analysis Tradi&onal*Hardware*Development:* – Incorporate micro & complex benchmarks throughout testbed evalua on process Benchmarks' Manual'' Topology'' • Common infrastructure for code reuse (simula on Hardware''à flight hardware) Specifica.ons' Component'' &'Netlist'' Performance'' Implementa.on' – Explore and compare several permuta ons in rapid succession Analysis' Crea.on' Results'
SpaceCubeX*Heterogeneous*Hardware*Development:*
Specifica.ons'
Applica.ons'' Compila.on'Environment' &'Benchmarks' Designer' Evolvable'Testbed' Inputs' Flight'Hardware' Architecture'' Performance'' Genera.on' Engineering'Hardware' Results' TRL* Hardware'Emula.on'
Design'Explora.on'' Simula.on' Feedback'Loop'
9 Simula on Testbed
User*Selec'on** Benchmarks*Suite* • Microkernels' • Applica:ons'
Simula'on*Environment:* ArchGen* Compile*&** • Processor' Mul'Gcore* FPGA* DSP* • Accelerator' Par''on* Sim*IF* Sim*IF* Sim*IF* Board' C* • Topology' Model' O* • Interconnects' N* Sim*IF* Sim*IF* (xml)' T* Simula'on* • Interfaces' Memory* I/O* • Memory' R* Interface* Simula'on** O* Translator* • Sensor' L* interfaces' Generator* Sim*IF* Sim*IF* • I/O' Sensor* Peripherals*
MONITOR*
Report* Component' Performance' @Power' Database' Models' @Area' @Radia:on' @Performance'
SpaceCubeX extends tradi onal simula on to incorporate en re hybrid system
10 Emula on Testbed
Benchmarks)Suite) • Microkernels+ • Applica/ons+ • System+
Emula&on)Environment:)
Compile)&)) Mul&Fcore)Board) FPGA)Board) DSP)Board) Par&&on) Emu)IF) Emu)IF) Emu)IF) Board+ C) Model+ O) (xml)+ N) Emu)IF) Emu)IF) T) Interface) R) Memory) I/O) Emula&on) O) Emulator) Generator) L) Emu)IF) Emu)IF) Sensor)Emulator) Peripherals)
MONITOR) Performance+ Models+ Report) 9Power+ 9Area+ 9Radia/on+ 9Performance+
SpaceCubeX emphasizes end-to-end board level architecture emula on
11 SpaceCubeX - Redsharc
REconfigurable Data-stream Hardware/So ware ARChitecture: • Provides rapid construc on of Mul -System-on-Chip pla orms (HW & SW) • Enables design space explora on of heterogeneous resources (Dev. Spirals) • System infrastructure supplied as IP, raising focus to algorithm development • Combines situa onal awareness with high performance processing – On-chip run- me control – Dynamic resource management – System and kernel-level performance tuning • Redsharc is not a high-level synthesis tool – Supports use of HLS tools use to create HW cores – Fills gaps in HLS tools: Integra on, communica on, and run- me management
Redsharc Philosophy: Interoperate and leverage current and emerging devices, tools, and technology to make developers more produc ve
12 Redsharc Development Spectrum
Accelera ng Development Maturity
Host PC: SW FPGA: SW FPGA: HW FPGA: HW FPGA: HW
On-Chip Kernel System System Towards Full-System Integra on Simula on Simula on Run- me
Redsharc infrastructure supports rapid tes ng across a variety of development environments to allow users to leverage preferred design and debugging tools such as Eclipse, GDB, ModelSim, and ChipScope.
13 Redsharc APIs
• Applica on Programming Interfaces: – So ware Kernel Interface – Hardware Kernel Interface – Dataflow Graph (DFG) – System – Scheduling • Run- me Execu on – Control Kernel Opera on
End-user leverages higher levels of abstrac on to implement applica on on heterogeneous resources
14 Redsharc Build Environment
DSP M.C.
Redsharc Generate flow constructs finished system from user provided design files
15 So ware Compila on Flow
Applica'ons+&+Benchmarks+
Task+ Task+ Task+
Task+
Manual+Task+Par''oning+ SoLware/Hardware+ CoBDesign+Library:! Hw+ Sw+ Sw+ Task+ Task+ Task+ ,!Communica4ons! ,!Interfaces! Hw+ Task+ ,!Protocols! ,!Run,4me!controls! Redsharc+Libraries+
Mul'Bcore+ Accelerator+ FPGA++ FPGA+Resource++ Library:+ Compiler+ Compiler+ Synthesis+ B!Interfaces! B!Interconnects! , Accelerator!! Cores! Bitstream+ Bitstream+ Executable+ Executable+ Executable+ Executable+ Core+0+ Core+1+ …! Core+n" DSP" FPGA+0" …! FPGA+n"
Evolvable+Testbed+
16 List of Benchmarks
Benchmark Description Micro Benchmarks Kernels to benchmark architecture subcomponents and measure system viability. ALHAT Analyze landing sites; Image processing (interpolation, correlation, kalman filter), Autonomous Landing and Hazard Avoidance and Detection System 4 sensors, task/data parallelism. ASPEN Artificial intelligence, multi-threaded, low level scheduler to satisfy constrained Automated Scheduling and Planning Environment mission goals. ROCKSTER Autonomous spacecraft tasking, geological feature identification, analysis, and Rock Segmentation Through Edge Regrouping data handling. Global, continuous real-time ocean radiometer for carbon assessment enabling ACE Mission Scenario research and coordinated activities based on live ocean color data; multi-spectral Aerosol-Cloud-Ecosystem image, data parallel processing Synthetic aperture radar and multi-beam lidar instruments produce voluminous raw data (future data rate concepts >1012 bits/second); benchmarks perform Radar/Lidar Mission Scenario terrain, vegetation and/or glacial ice mapping; high data rate processing, data parallelism. Hyper-spectral Earth science event detection algorithms (fire, flood, coastal HyspIRI Mission Scenario spills); Autonomously reconfigures for Visible to Short Wavelength Infrared data for real-time mapping and transmission.
17 Result Metrics
Parameter Units Power Watts Estimated board size cm2 Estimated board weight kg Processing performance Instructions or Operations per second Input Output Bandwidth Bits per second Radiation Tolerance Total Ionizing Dose and Single Event Upset rate Cost $
Provide designers and scien sts more feedback earlier in development process
18 Example Results
Performance unit (Area, Architecture 1 Power, Processing…) Architecture 2
Architecture 3
Single CPU baseline 1 1 1 N N 1 N N App App App App App App App App Sequential Streaming Server Bit level
19 Ge ng “On-Board”
• We want to learn from the community – Understand applica ons and instruments – Advice on applica on/algorithm implementa ons – Sugges ons on devices to consider • How is the current technology performing? • Hard restric ons on data precision? • Applica on’s demand for radia on tolerance?
20 Summary
SpaceCubeX aims to: • Alleviate computa on limita ons for on-board processing • Enable selec on of SWAP efficient processing architecture: – Impact on mission capabili es, cost, and risk • Reduce risk due to supply chain disrup on • Require minimal extensions to integrate new devices • Support persistent, long-term mission opera ons • Provide poten al infusion into many missions: – (P)ACE, HyspIRI, GEO-CAPE, ICESat-2…
The SpaceCubeX Project is a structured approach to assess and develop hybrid architectures, fundamentally changing avionics processing architecture development process.
21 QUESTIONS
22 BACKUP SLIDES
23 Project Milestones
• Project start 05/15 • Complete ini al scenario benchmarks 11/15 • Ini al architecture 12/15 • Final architecture 03/16 • Final benchmark package 04/16 • Ini al emula on testbed 12/16 • Final emula on testbed 03/17 • Benchmark demonstra on 04/17
24 Development Abstrac ons
Hardware Kernel Interface So ware Kernel Interface
Data accesses are abstracted from the kernel implementa on
25 TRL Assessment
Category Component Initial 6 Mos 12 Mos 18 Mos Final Component Selection 3 3 4 4 5 Hardware Interconnect Topology 3 3 4 4 5 Architecture Control Software 3 3 4 4 5 Hardware Architecture Subtotal 3 3 4 4 5 Programming Model 4 5 5 5 5 Total 3 3 4 4 5
26