Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Lecture 18: Introduction to Multiprocessors Prepared and presented by: Kurt Keutzer with thanks for materials from Kunle Olukotun, Stanford; David Patterson, UC Berkeley 1 Why Multiprocessors? Needs Relentless demand for higher performance » Servers » Networks Commercial desire for product differentiation Opportunities Silicon capability Ubiquitous computers 2 Exploiting (Program) Parallelism Levels of Parallelism Process Thread Loop Instruction 1 10 100 1K 10K 100K 1M Grain Size (instructions) 3 Exploiting (Program) Parallelism -2 Levels of Parallelism Process Thread Loop Instruction Bit 1 10 100 1K 10K 100K 1M Grain Size (instructions) 4 Need for Parallel Computing Diminishing returns from ILP Performance » Limited ILP in programs » ILP increasingly expensive to exploit Peak performance increases linearly with more processors » Amhdahl’s law applies Adding processors is inexpensive » But most people add memory also 2P+M Performance Die Area 2P+2M P+M Die Area 5 What to do with a billion transistors ? 1 clk Technology changes the cost and performance of computer elements in a non-uniform manner » logic and arithmetic is becoming plentiful and cheap » wires are becoming slow and scarce This changes the tradeoffs between alternative architectures » superscalar doesn’t scale well – global control and data So what will the architectures of the future be? 1998 2001 2004 64 x the area 4x the speed slower wires 2007 3 (10, 16, 20?) clks 6 Elements of a multiprocessing system General purpose/special purpose Granularity - capability of a basic module Topology - interconnection/communication geometry Nature of coupling - loose to tight Control-data mechanisms Task allocation and routing methodology Reconfigurable » Computation » Interconnect Programmer’s model/Language support/ models of computation Implementation - IC, Board, Multiboard, Networked Performance measures and objectives [After E. V. Krishnamurty - Chapter 5 7 Use, Granularity General purpose attempting to improve general purpose computation (e.g. Spec benchmarks) by means of multiprocessing Special purpose attempting to improve a specific application or class of applications by means of multiprocessing Granularity - scope and capability of a processing element (PE) Nand gate ALU with registers Execution unit with local memory RISC R1000 processor 8 Topology Topology - method of interconnection of processors Bus Full-crossbar switch Mesh N-cube Torus Perfect shuffle, m-shuffle Cube-connected components Fat-trees 9 Coupling Relationship of communication among processors Shared clock (Pipelined) Shared registers (VLIW) Shared memory (SMM) Shared network 10 Control/Data Way in which data and control are organized Control - how the instruction stream is managed (e.g. sequential instruction fetch) Data - how the data is accessed (e.g. numbered memory addresses) Multithreaded control flow - explicit constructs: fork, join, wait, control program flow - central controller Dataflow model - instructions execute as soon as operands are ready, program structures flow of data, decentralized control 11 Task allocation and routing Way in which tasks are scheduled and managed Static - allocation of tasks onto processing elements predetermined before runtime Dynamic - hardware/software support allocation of tasks to processors at runtime 12 Reconfiguration Computational restructuring of computational elements » reconfigurable - reconfiguration at compile time » dynamically reconfigurable- restructuring of computational elements at runtime Interconnection scheme switching network - software controlled reconfigurable fabric 13 Programmer’s model How is parallelism expressed by the user? Expressive power Process-level parallelism » Shared-memory » Message-passing Operator-level parallelism Bit-level parallelism Formal guarantees Deadlock-free Livelock free Support for other real-time notions Exception handling 14 Parallel Programming Models Message Passing » Fork thread – Typically one per node » Explicit communication – Send messages – send(tid, tag, message) – receive(tid, tag, message) » Synchronization – Block on messages (implicit sync) – Barriers Shared Memory (address space) » Fork thread – Typically one per node » Implicit communication – Using shared address space – Loads and stores » Synchronization – Atomic memory operators – barriers 15 Message Passing Multicomputers Computers (nodes) connected by a network » Fast network interface – Send, receive, barrier » Nodes not different than regular PC or workstation Cluster conventional workstations or PCs with fast network » cluster computing » Berkley NOW » IBM SP2 Node P P P M M M Network 16 Shared-Memory Multiprocessors Several processors share one address space » conceptually a shared memory » often implemented just like a multicomputer – address space distributed over private memories Communication is implicit » read and write accesses to shared memory locations Synchronization » via shared memory locations – spin waiting for non-zero » barriers P P P Network M Conceptual Model P P P M M M Network Actual Implementation 17 Cache Coherence - A Quick Overview With caches, action is required to prevent access to stale data P1 P2 PN » Processor 1 may read old data from its cache instead of new data in memory or » Processor 3 may read old data from memory rather than new data in Processor 2’s cache $ $ $ Solutions » no caching of shared data – Cray T3D, T3E, IBM RP3, BBN Butterfly » cache coherence protocol – keep track of copies – notify (update or invalidate) on writes Network M A:3 P1: Rd(A) P2: P3: Rd(A) Wr(A,5) Rd(A) 18 Implementation issues Underlying hardware implementation Bit-slice Board assembly Integration in an integrated-circuit Exploitation of new technologies DRAM integration on IC Low-swing chip-level interconnect 19 Performance objectives Objectives Speed Power Cost Ease of programming/time to market/ time to money In-field flexibility Methods of measurement Modeling Emulation Simulation » Transaction » Instruction-set » Hardware 20 Flynn’s Taxonomy of Multiprocessing Single-instruction single-datastream (SISD) machines Single-instruction multiple-datastream (SIMD) machines Multiple-instruction single-datastream (MISD) machines Multiple-instruction multiple-datastream (MIMD) machines Examples? 21 Examples Single-instruction single-datastream (SISD) machines » Non-pipelined Uniprocessors Single-instruction multiple-datastream (SIMD) machines » Vector processors (VIRAM) Multiple-instruction single-datastream (MISD) machines » Network processors (Intel IXP1200 Multiple-instruction multiple-datastream (MIMD) machines » Network of workstations (NOW) 22 Predominant Approaches Pipelining ubiquitious Much academic research focused on performance improvements of ``dusty decks’’ Illiac 4 - Speed-up of Fortran SUIF, Flash - Speed-up of C Niche market in high-performance computing Cray Commercial support for high-end servers Shared-memory multiprocessors for server market Commercial exploitation of silicon capability General purpose: Super-scalar, VLIW Special purpose: VLIW for DSP, Media processors, Network processors Reconfigurable computing 23 C62x Pipeline Operation Pipeline Phases Fetch Decode Execute PG PS PW PR DP DC E1 E2 E3 E4 E5 Single-Cycle Throughput Operate in Lock Step Fetch » » » » PG PS PW PR Decode » » DP DC Instruction Dispatch Instruction Decode Execute » E1 - E5 Execute 1 through Execute 5 E3 E2 E1 DC DP PR PW E5 E4 E3 E2 E1 DC DP Program Address Generate Program Address Send Program Access Ready Wait Program Fetch Packet Receive Execute Packet 1 PG PS PW PR DP DC Execute Packet 2 PG PS PW PR DP Execute Packet 3 PG PS PW PR Execute Packet 4 PG PS PW Execute Packet 5 PG PS Execute Packet 6 PG Execute Packet 7 E1 DC DP PR PW PS PG E2 E1 DC DP PR PW PS E4 E3 E2 E1 DC DP PR E5 E4 E3 E2 E1 DC E5 E4 E3 E2 E1 E5 E4 E5 E3 E4 E5 E2 E3 E4 E5 24 Superscalar: PowerPC 604 and Pentium Pro Both In-order Issue, Out-of-order execution, In-order Commit 25 IA-64 aka EPIC aka VLIW Compiler schedules instructions Encodes dependencies explicitly » saves having the hardware repeatedly rediscover them Instruction Cache Instruction Issue Support speculation » speculative load » branch prediction Really need to make communication explicit too » still has global registers and global instruction issue Register File 26 Phillips Trimedia Processor 27 TMS320C6201 Revision 2 Program Cache / Program Memory 32-bit address, 256-Bit data512K Bits RAM Pwr Dwn C6201 CPU Megamodule Program Fetch Control Registers Instruction Dispatch Host Port Interface 4-DMA Instruction Decode Data Path 1 Data Path 2 A Register File Control Logic B Register File Test Emulation Ext. Memory Interface L1 S1 M1 D1 D2 M2 S2 L2 Interrupts 2 Timers Data Memory 32-Bit address, 8-, 16-, 32-Bit data 512K Bits RAM 2 Multichannel buffered serial ports (T1/E1) 28 TMS320C6701 DSP Block Diagram Program Cache/Program Memory 32-bit address, 256-Bit data 512K Bits RAM Power ’C67x Floating-Point CPU Core Down Program Fetch Host Port Interface Control Registers Instruction Dispatch 4 Channel DMA Instruction Decode Data Path 1 Data Path 2 A Register File Control Logic B Register File Test Emulation L1 S1 M1 D1 External Memory Interface D2 M2 S2 L2 Interrupts 2 Timers Data Memory 32-Bit address 8-, 16-, 32-Bit data 512K Bits RAM 2 Multichannel buffered serial ports (T1/E1) 29 TMS320C67x CPU Core ’C67x Floating-Point CPU Core Program Fetch Instruction Dispatch Control Registers Instruction Decode Data Path 1 Data Path 2 A Register File B Register File Control Logic Test Emulation L1 S1 M1 D1 Arithmetic Logic Unit Auxiliary Logic Unit D2 M2 S2 L2 Multiplier Unit Interrupts Floating-Point Capabilities 30 Single-Chip Multiprocessors CMP Build a multiprocessor on a single chip » linear increase in peak performance » advantage of fast interaction between processors Fine grain threads » make communication and synchronization very fast (1 cycle) » break the problem into smaller pieces P P P P $ $ $ $ $ M Memory bandwidth » Makes more effective use of limited memory bandwidth Programming model » Need parallel programs 31 Intel IXP1200 Network Processor SDRAM Ctrl » » » MicroEng PCI Interface ICache SA Core MicroEng MicroEng Hash Engine IX Bus Interface MicroEng DCache Mini DCache MicroEng 6 micro-engines IX Bus Interface » » less critical tasks Hash engine » packet I/O connect IXPs – scalable StrongARM » Scratch Pad SRAM RISC engines 4 contexts/eng 24 threads total level 2 lookups PCI interface MicroEng SRAM Ctrl 32 IXP1200 MicroEngine from SRAM 32 SRAM Read XFER Registers to SRAM 32 SRAM Write XFER Registers 64 GPRs (A-Bank) ALU 32 SDRAM Read XFER Registers 64 GPRs (B-Bank) from SDRAM to SDRAM 32-bit RISC instruction set Multithreading support for 4 threads » 32 SDRAM Read XFER Registers Maximum switching overhead of 1 cycle 128 32-bit GPRs in two banks of 64 Programmable 1KB instruction store (not shown in diagram) 128 32-bit transfer registers Command bus arbiter and FIFO (not shown in diagram) 33 IXP1200 Instruction Set Powerful ALU instructions: » can manipulate word and part of word quite effectively Swap-thread on memory reference » Hides memory latency » sram[read, r0, base1, offset, 1], ctx_swap Can use an “intelligent” DMA-like controller to copy packets to/from memory » sdram[t_fifo_wr, --, pkt_bffr, offset, 8] Exposed branch behavior » can fill variable branch slots » can select a static prediction on a per-branch basis ARM mov r1, r0, lsl #16 mov r1, r1, r0, asr #16 add r0, r1, r0, asr #16 IXP1200 ld_field_w_clr[temp, 1100, accum] alu_shf[accum, temp, +, accum, <<16] 34 UCB: Processor with DRAM (PIM) IRAM, VIRAM Put the processor and the main memory on a single chip » much lower memory latency » much higher memory bandwidth P V M But » need to build systems with more than one chip 64Mb SDRAM Chip Internal - 128 512K subarrays 4 bits per subarray each 10ns 51.2 Gb/s External - 8 bits at 10ns, 800Mb/s 1 Integer processor ~ 100KBytes DRAM 1 FP processor ~ 500KBytes DRAM 1 Vector Unit ~ 1 MByte DRAM 35 IRAM Vision Statement Microprocessor & DRAM on a single chip: » on-chip memory latency 5-10X, bandwidth 50-100X » improve energy efficiency 2X-4X (no off-chip bus) » serial I/O 5-10X v. buses » smaller board area/volume » adjustable memory size/width I/O I/O Bus Proc $ $ L2$ Bus L o f ga i b c D R A M I/O I/O Proc Bus D R A M D Rf Aa Mb 36 Potential Multimedia Architecture “New” model: VSIW=Very Short Instruction Word! » » » » » Compact: Describe N operations with 1 short instruct. Predictable (real-time) performance vs. statistical performance (cache) Multimedia ready: choose N*64b, 2N*32b, 4N*16b Easy to get high performance Compiler technology already developed, for sale! – Don’t have to write all programs in assembly language 37 Revive Vector (= VSIW) Architecture! Cost: ≈ $1M each? Low latency, high BW memory system? Code density? Compilers? Performance? Power/Energy? Limited to scientific applications? Single-chip CMOS MPU/IRAM IRAM Much smaller than VLIW For sale, mature (>20 years) Easy scale speed with technology Parallel to save energy, keep perf Multimedia apps vectorizable too: N*64b, 2N*32b, 4N*16b 38 V-IRAM1: 0.18 µm, Fast Logic, 200 MHz 1.6 GFLOPS(64b)/6.4 GOPS(16b)/16MB 4 x 64 or 8 x 32 or 16 x 16 + x 2-way Superscalar Processor I/O Vector Instruction Queue I/O ÷ Load/Store Vector Registers 16K I cache 16K D cache 4 x 64 4 x 64 Serial I/O Memory Crossbar Switch M I/O M 4…x 64 I/O M M M M M M M … M 4…x 64 M x 64 … 4… … M M M M M M M M M 4…x 64 M M M M M … M 4… x 64 … M M M M … 39 Tentative VIRAM-1 Floorplan 0.18 µm DRAM 16-32 MB in 16 banks x Memory (128 Mbits / 16 MBytes) 256b 0.18 µm, 5 Metal Logic ≈ 200 MHz MIPS IV, C 16K I$, 16K D$ RingP ≈ 4 200 MHz based 4 Vector Pipes/Lanes U I/O FP/int. vector units Switch +$ die: ≈ 20x20 mm xtors: ≈ 130-250M power: ≈2 Watts Memory (128 Mbits / 16 MBytes) 40 Tentative VIRAM-”0.25” Floorplan Memory (32 Mb / 4 MB) C P 1 VU U +$ Memory (32 Mb / 4 MB) Demonstrate scalability via 2nd layout (automatic from 1st) 8 MB in 2 banks x 256b, 32 subbanks ≈ 200 MHz CPU, 8K I$, 8K D$ 1 ≈ 200 MHz FP/int. vector units die: ≈ 5 x 20 mm xtors: ≈ 70M power: ≈0.5 Watts Kernel Comp. iDCT Clr.Conv. Convol. FP Matrix V-1 6.40 3.10 2.95 3.16 3.19 GOPS V-0.25 1.6 0.8 0.8 0.8 0.8 41 Stanford: Hydra Design Centralized Bus Arbitration Mechanisms CPU 0 L1 Ins t. Cac he CP2 CPU 1 L1 Data Cache & Spe cula tion Bits L1 Ins t. Cac he CP2 L1 Data Cache & Spe cula tion Bits CPU 1 Me mory Controller CPU 0 Me mory Controller CP2 CPU 2 L1 Ins t. Cac he L1 Data Cache & Spe cula tion Bits CPU 2 Me mory Controller CPU 3 L1 Ins t. Cac he CP2 L1 Data Cache & Spe cula tion Bits CPU 3 Me mory Controller Write-through Bus (64b) Read/Replace Bus (256b) Speculation Write Buffers #0 #1 #2 #3 On-chip L2 Cache retire Rambus Memory Interface I/O Bus Interface DRAM Main Memory I/O Devices Single-chip multiprocessor Four processors Separate primary caches Write-through data caches to maintain coherence Shared 2nd-level cache Separate read and write busses Data Speculation Support 42 Mescal Architecture Scott Weber University of California at Berkeley 43 Outline Architecture rationale and motivation Architecture goals Architecture template Processing elements Multiprocessor architecture Communication architecture 44 Architectural Rationale and Motivation Configurable processors have shown orders of magnitude performance improvements Tensilica has shown ~2x to ~50x performance improvements » Specialized functional units » Memory configurations Tensilica matches the architecture with software development tools Memory RegFile FU ICache Set memory parameters Add DCT and Huffman blocks for a JPEG app Configuration Memory RegFile FU FU DCT HUF FU ICache 45 Architectural Rationale and Motivation In order to continue this performance improvement trend » Architectural features which exploit more concurrency are required » Heterogeneous configurations need to be made possible » Software development tools support new configuration options Memory RegFile FU FU DCTHUF FU ...begins to look like a VLIW... PE Memory PE PE PE RegFile PE PE FU FU DCTHUF FU PE ICache PE PE PE PE PE PE PE PE PE ...concurrent processes are required in order to continue performance improvement trend... PE PE ICache ...generic mesh PE may not suit the application’s PE topology... PE PE PE PE PE ...configurable VLIW PEs and network topology... PE 46 Architecture Goals Provide template for the exploration of a range of architectures Retarget compiler and simulator to the architecture Enable compiler to exploit the architecture Concurrency » Multiple instructions per processing element » Multiple threads per and across processing elements » Multiple processes per and across processing elements Support for efficient computation » Special-purpose functional units, intelligent memory, processing elements Support for efficient communication » Configurable network topology » Combined shared memory and message passing 47 Architecture Template Prototyping template for array of processing elements » Configure processing element for efficient computation » Configure memory elements for efficient retiming » Configure the network topology for efficient communication Memory RegFile ...configure PE... Memory RegFile FU FU FU FU FU FU FU DCTHUF FU ICache ICache ...configure memory elements... Memory Memory RegFile RegFile FU FU DCTHUF FU ICache ...configure PEs and network to match the application... 48 Range of Architectures Memory System Register File FU Instruction Cache Scalar Configuration EPIC Configuration EPIC with special FUs Mesh of HPL-PD PEs Customized PEs, network Supports a family of architectures » Plan to extend the family with the micro-architectural features presented 49 Range of Architectures PE PE Memory System PE Register File PE PE PE FU PE FU PE FU FU PE Instruction Cache FU Scalar Configuration EPIC Configuration EPIC with special FUs Mesh of HPL-PD PEs Customized PEs, network Supports a family of architectures » Plan to extend the family with the micro-architectural features presented 50 Range of Architectures Memory System Register File FU FU DES DCT FFT Instruction Cache Scalar Configuration EPIC Configuration EPIC with special FUs Mesh of HPL-PD PEs Customized PEs, network Supports a family of architectures » Plan to extend the family with the micro-architectural features presented 51 Range of Architectures PE Memory PE System PE PE Register File PE PE FU PE FU DES DCT FFT PE Instruction Cache PE Scalar Configuration EPIC Configuration EPIC with special FUs Mesh of HPL-PD PEs Customized PEs, network Supports a family of architectures » Plan to extend the family with the micro-architectural features presented 52 Range of Architectures PE PE PE PE PE PE PE PE Scalar Configuration EPIC Configuration EPIC with special FUs Mesh of HPL-PD PEs Customized PEs, network Supports a family of architectures » Plan to extend the family with the micro-architectural features presented 53 Range of Architectures (Future) SDRAM Ctrl MicroEng PCI Interface ICache SA Core MicroEng MicroEng Hash Engine IX Bus Interface MicroEng DCache Mini DCache MicroEng Template support for such an architecture Prototype architecture Software development tools generated » Generate compiler » Generate simulator Scratch Pad SRAM MicroEng SRAM Ctrl IXP1200 Network Processor (Intel) 54 The RAW Architecture Slides prepared by Manish Vachhrajani 55 Outline RAW architecture » Overview » Features » Benefits and Disadvantages Compiling for RAW » » » » Overview Structure of the compiler Basic block compilation Other techniques 56 RAW Machine Overview Scalable architecture without global interconnect Constructed from Replicated Tiles » Each tile has a mP and a switch » Interconnect via a static and dynamic network 57 RAW Tiles Simple 5 stage pipelined mP w/ local PC(MIMD) » Can contain configurable logic Per Tile IMEM and DMEM, unlike other modern architectures mP contains ins. to send and recv. data IMEM DMEM PC REGS CL PC SMEM Switch 58 RAW Tiles(cont.) Tiles have local switches » Implemented with a stripped down mP » Static Network – Fast, easy to implement – Need to know data transfers, source and destintation at compile time » Dynamic Network – Much slower and more complex – Allows for messages whose route is not known at compile time 59 Configurable Hardware in RAW Each tile Contains its own configurable hardware Each tile has several ALUs and logic gates that can operate at bit/byte/word levels Configurable interconnect to wire componenets together Coarser than FPGA based implementations 60 Benefits of RAW Scalable » Each tile is simple and replicated » No global wiring, so it will scale even if wire delay doesn’t » Short wires and simple tiles allow higher clock rates Can target many forms of Parallelism Ease of design » Replication reduces design overhead » Tiles are relatively simple designs » simplicity makes verification easier 61 Disadvantages of RAW Complex Compilation » Full space-time compilation » Distributed memory system » Need sophisticated memory analysis to resolve “static references” Software Complexity » Low-level code is complex and difficult to examine and write by hand Code Size? 62 Traditional Operations on RAW How does one exploit the Raw architecture across function calls, especially in libraries? » Can we easily maintain portability with different tile counts? Memory Protection and OS Services » Context switch overhead » Load on dynamic network for memory protection and virtual memory? 63 Compiling for RAW machines Determine available parallelism Determine placement of memory items Discover memory constraints » Dependencies between parallel threads » Disambiguate memory references to allow for static access to data elements » Trade-off memory dependence and Parallelism 64 Compiling for RAW(cont.) Generate route instructions for switches » static network only Generate message handlers for dynamic events » Speculative execution » Unpredictable memory references Optimal partitioning algorithm is NP complete 65 Structure of RAWCC Source Language Traditional Dataflow Optimizations Build CFG Partition data to increase static accesses Partition instructions to allow parallel execution allocate data to tiles to minimize communication overhead MAPS System Space-time scheduler RAW executable 66 The MAPS System Manages memory to generate static promotions of data structures For loop accesses to arrays uses modulo unrolling For data structures, uses SPAN analysis package to identify potential references and partition memory » structures can be split across processing elements. 67 Space-Time Scheduler For Basic Blocks » » » » Maps instructions to processors Maps scalar data to processors Generates communication instructions Schedules computation and communication For overall CFG, performs control localization 68 Basic Block Orchestrator Initial Code Transformation Instruction Partitioner Global Data Partitioner Data & Ins. Placer Comm Code Generator Event Scheduler All values are copied to the tiles that work on the data from the home tile Within a Block, all access are local At the end of a block, values are copied to home tiles 69 Initial Code Transformation Convert Block to static single assignment form » removes false dependencies » Analagous to register renaming Live on entry, and live on exit variables marked with dummy instructions » Allows for overlap of “stitch” code with useful work 70 Instruction Partitioner Partitions stream into multiple streams, one for each tile Clustering » Partition instructions to minimize runtime considering only communication Merging » Reduces cluster count to match tile count » Uses a heuristic based algorithm to achieve good balance and low communication overhead 71 Global Data Partitioner Partitions global data for assignment to home locations » Local data is copied at the start of a basic block Summarize instruction stream’s data access pattern with affinity Maps instructions and data to virtual processors » Map instructions, optimally place data based on affinity » Remap instructions with data placement knowledge » Repeat until local minima is reached Only real data are mapped, not dummies formed in ICT 72 Data and Instruction Placer Places data items onto physical tiles » driven by static data items Places instructions onto tiles » Uses data information to determine cost Takes into account actual model of communications network Uses a swap based greedy allocation 73 Event Scheduler Schedules routing instructions as well as computation instructions in a basic block Schedules instructions using a greedy list based scheduler Switch schedule is ensured to be deadlock free » Allows tolerance of dynamic events 74 Control Flow Control Localization » Certain branches are enveloped in macro instructions, and the surrounding blocks merged » Allows branch to occur only on one tile Global Branching » Done through target broadcast and local branching 75 Performance RAW achieves anywhere from 1.5 to 9 times speedup depending on application and tile count Applications tested were particularly well suited to RAW Heavily dependent integer programs may do poorly(encryption, etc.)) Depends on its ability to statically schedule and localize memory accesses 76 Future Work Use multisequential execution to run multiple applications simultaneously » Allow static communication between threads known at compile time » Minimize dynamic overhead otherwise Target ILP across branches more agressively Explore configurability vs. parallelism in RAW 77 Reconfigurable processors Adapt the processor to the application » » Builds on FPGA technology » » » special function units special wiring between function units FPGAs are inefficient – a multiplier built from an FPGA is about 100x larger and 10x slower than a custom multiplier. Need to raise the granularity – configure ALUs, or whole processors Memory and communication are usually the bottleneck – not addressed by configuring a lot of ALUs Programming model » » Difficult to program Verilog 78 SCORE Stream Computation Organized for Reconfigurable Execution Eylon Caspi Michael Chu André DeHon Randy Huang Joseph Yeh John Wawrzynek Nicholas Weaver 79 Opportunity High-throughput, regular operations can be mapped spatially onto FPGA-like (programmable, spatial compute substrate) achieving higher performance » (throughput per unit area) than conventional, programmable devices » (e.g. processors) 80 Problem Only have raw devices Solutions non-portable Solutions not scale to new hardware Device resources exposed to developer Little or no abstraction of implementations Composition of subcomponents hard/ad hoc No unifying computational model or run-time environment 81 Introduce: SCORE Compute Model » » » » » virtualizes RC hardware resources supports automatic scaling supports dynamic program requirements efficiently provides compositional semantics defines runtime environment for programs 82 Viewpoint SCORE (or something like it) is a necessary condition to enable automatic exploitation of new RC hardware as it becomes available. Automatic exploitation is essential to making RC a long-term viable computing solution. 83 Outline Opportunity Problem Review » related work » enabling hardware Model » execution » programmer Preliminary Results Challenges and Questions ahead 84 …borrows heavily from... RC, RTR P+FPGA Dataflow Streaming Dataflow Multiprocessors Operating System (see working paper) Tried to steal all the good ideas :-) build a coherent model exploit strengths of RC 85 Enabling Hardware High-speed, computational arrays » [250MHz, HSRA, FPGA’99] Large, on-chip memories » [2Mbit, VLSI Symp. ‘99] » [allow microsecond reconfiguration] Processor and FPGA hybrids » [GARP, NAPA, Triscend, etc.] 86 BRASS Architecture 87 Array Model 88 Platform Vision Hardware capacity scales up with each generation » Faster devices » More computation » More memory With SCORE, old programs should run on new hardware » and exploit the additional capacity automatically 89 Example: SCORE Execution 90 Spatial Implementation 91 Serial Implementation 92 Summary: Elements of a multiprocessing system General purpose/special purpose Granularity - capability of a basic module Topology - interconnection/communication geometry Nature of coupling - loose to tight Control-data mechanisms Task allocation and routing methodology Reconfigurable » Computation » Interconnect Programmer’s model/Language support/ models of computation Implementation - IC, Board, Multiboard, Networked Performance measures and objectives [After E. V. Krishnamurty - Chapter 5 93 Conclusions Portions of multi/parallel processing have become successful » Pipelining ubiquitious » Superscalar ubiquitious » VLIW successful in DSP, Multimedia - GPP? Silicon capability re-invigorating multiprocessor research » GPP - Flash, Hydra, RAW » SPP - Intel IXP 1200, IRAM/VIRAM, Mescal Reconfigurable computing has found a niche in wireless communications Problem of programming models, languages, computational models etc. for multiprocessors still largely unsolved 94