Chap1 Introduction. Outline. An Example System. 1.1 Overview. Computer organization and architecture. Computer organization and architecture

Transcription

1 Computer Architecture Chap Introduction Zheng Qinghua CS Department of XJTU Outline Overview What s the Computer Architecture Classification of Computer Architecture Quantitative Design Principles Evaluation and Benchmark Overview Why study computer architecture? Design better programs, including system software such as compilers, operating systems, and device drivers. Optimize program behavior. Evaluate (benchmark) computer system performance. Understand the tradeoffs relationship among time, space, and price cost. An Example System What does it all mean? Computer organization and architecture Computer organization physical aspects of computer systems. E.g., circuit design, control signals, memory types. Key: How does a computer work? Computer architecture Logical aspects of system as seen by the assembly programmer or OS designer. E.g., instruction sets, instruction formats, data types, addressing modes, I/O mode. Key: How do we design a computer? Computer organization and architecture Principle of Equivalence of Hardware and Software: Anything that can be done with software can also be done with hardware, and anything that can be done with hardware can also be done with software.* At the most basic level, a computer is a device consisting of three pieces: A processor to interpret and execute programs A memory to store both data and programs A mechanism for * Assuming speed is not a concern. transferring data to and from the outside world, i.e. I/O system.

2 .2 What is Computer Architecture? Research Topics of Single Computer Architecture Definition[Amdahl, Blaaw, and Brooks, 964.]: the attributes of a [computing] system as seen by the programmer, i.e., the conceptual structure and functional behavior, as distinct from the organization of the data flows and controls the logic design, and the physical implementation. Remark: G.M.Amdahl was the general system designer of IBM,he proposed this conception in 964 when he designed the IBM 360 computer. Input/Output and Storage Memory Hierarchy Hard Disks, CD, NAS, SAN DRAM L2 Cache L Cache VLSI Instruction Set Architecture Pipelining, Hazard Resolution, Superscalar, Reordering, Prediction, Speculation, Vector, DSP Emerging Technologies Interleaving Memories Coherence, Bandwidth, Latency RAID Addressing, Protection, Exception Handling Pipelining and Instruction Level Parallelism Research Topics of Multiple Computer Architecture The Tasks of a Computer Designer P M P M P M P M. Shared Memory, 2. Message Passing, 3. Data Parallelism Implementation Complexity Evaluate Existing Systems for Bottlenecks S Interconnection Network Processor-Memory-Switch Multiprocessors Networks and Interconnections Network Interfaces Topologies, Routing, Bandwidth, Latency, Reliability Implement Next Generation System Workloads Technology Trends Simulate New Designs and Organizations Benchmarks Basic principle: balance Cost /Performance Computer Systems: Technology Trends Supercomputers Multi Core Processor Massively Parallel Processors Network of SMP Mini-supercomputers Workstations Minicomputers Mainframes Workstations Supercomputers PC s Cloud Computing Embedded Computers.3 The Computer Level Hierarchy Complex computer systems employ virtual machine layers. Each one is an abstraction of the level below it, execute their own particular instructions, calling upon machines at lower levels to perform tasks as required. Computer circuits ultimately carry out the work. 2

3 .3 The Computer Level Hierarchy Level 6: The User Level Program execution and user interface level. The level with which we are most familiar. Level 5: High-Level Language Level The level with which we interact when we write programs in languages such as C, Python, XML and Java. Level 4: Assembly Language Level Acts upon assembly language produced from Level 5, as well as instructions programmed directly at this level..3 The Computer Level Hierarchy Level 3: System Software Level Controls and Schedule executing processes on the system. Protects system resources. Level 2: Machine Level Also known as the Instruction Set Architecture (ISA) Level. Consists of instructions that are particular to the architecture of the machine. Programs written in machine language need no compilers, interpreters, or assemblers..3 The Computer Level Hierarchy Level : Control Level A control unit decodes and executes instructions and moves data through the system. Control units can be microprogrammed or hardwired. A microprogram is a program written in a low-level language that is implemented by the hardware. Hardwired control units consist of hardware that directly executes machine instructions. Level 0: Digital Logic Level This level is where we find digital circuits (the chips). Digital circuits consist of gates and wires. These components implement the mathematical logic of all other levels. )CPUs Single CPU Multiple Processors Parallel Processing Correlated Processing Superscalar Pipeline Symmetric Multiple Processors MPP (Massively Parallel Processors) Cluster, Cloud 2)Flynn Categories(996) () SISD :Single Instruction Single Data stream. (2) SIMD:Single Instruction Stream Multiple Data stream. (3) MISD:Single Data stream for Multiple Instructions. (4) MIMD:Multiple Instruction stream Multiple Data stream..4 Classification SISD:Single Instruction Single Data Stream SIMD:Single Instruction Multiple Data stream Ex:PDP- IBM-360/370 PC 8086 Z-80 Ex:Vectors Operate(+,-,*) ILLAC Ⅳ(64 Units) Array Machine Pentium s padd [b,w,d] 3

4 SIMD Computers MISD:Multiple Instructions Single Data Stream Such computers require multiple processing elements. Can share global memory among PEs or each PE might have local memory Control Unit controls signals to all PEs (one instruction is sent out, each PE has its own memory to work on) SIMD Computer characteristics: distribute processing over large amount of hardware, operate concurrently on many different data elements, perform same computation on each data element Commonly used for vector applications. MISD is a controversy architecture, no practical application. MISD Computers Logically, such a computer would issue several instructions simultaneously on the same datum For instance, perform six different analysis of the information stored in a single array Such an operation can be carried out on a SISD computer as sequential analysis, or on a MIMD computer MISD computers would be very restrictive in terms of useful applications and are therefore not constructed MIMD : Multiple Instructions Multiple Data Streams EX:IBM 308/3084 Univac 00/80,Cray-2 etc. MIMD Computers These are true parallel processing computers These computers have many independent processors (a computer with two somewhat dependent processors is not a MIMD) MIMD Characteristics: distribute processing over independent processors, share resources (sometimes including main memory), processors operate independently and concurrently, each processor can also run its own program Two extremes for MIMD computers: global memory (tightly coupled) local memory (loosely coupled).5 Quantitative Principles for Computer Architecture Designing Make the Common Case Fast Amdahl s Law CPU Performance Equation, decrease CPI Clock cycle time CPI: Cycles Per Instruction IPC: Instruction Count per Cycle Principles of Locality Take advantage of Parallelism 4

5 Principle:Make the common case fast. E.g.: Cache,short codec for common use instruction, Amdahl s Law: Relates total speedup of a system to the speedup of some portion of that system. Amdal Law ExTime new = ExTime old x ( - Fraction enhanced ) + Fraction enhanced Speedup enhanced Speedup overall = ExTime old ExTime new = ( - Fraction enhanced ) + Fraction enhanced Speedup enhanced Example -: 例 : 设某计算机系统执行程序中, 可向量化部分最大可达 70% 一种实现方法是采用阵列多处理机即硬件方式实现, 使 60% 的向量化指令执行速度加快到原来的 0 倍, 另一种方法是采用优化编译系统的方法实现, 使向量化程序可达 70%, 且速度可增加 3 倍 请比较这两种方案哪种更优 S硬件 ( 60%) 0.6 / S编译 ( 70%) 0.7 / 采用硬件的方法且有更高的加速比 年 3 月 If Amdahl Law can be used in parallel computer system or multi-processors to find the maximum expected improvement to an overall system when only part of the system is improved? References: Think it over Principle2: CPU Performance Equation CPU time ( Seconds ) = Instructions x Cycles x Seconds Program Program Instruction Cycle Inst Count CPI Clock Rate Program X Compiler X (X) Inst. Set. X X Organization X X Principle2: Lower Cycles Per Instruction (CPI) CPI = (CPU Time * Clock Rate) / Instruction Count = Cycles / Instruction Count CPU_ Time Cycle _ Time* Instruction Frequency CPI n i CPI i * F i where n i CPI i * I i F i Number of instructions of type I. Ii Instruction _ Count Technology X Invest Resources where time is Spent! 5

6 Principle2: Lower Cycles Per Instruction Suppose we have a machine where we can count the frequency with which instructions are executed. We also know how many cycles it takes for each instruction type. Base Machine (Reg / Reg) Op Freq Cycles CPI(i) (% Time) ALU 50%.5 (33%) Load 20% 2.4 (27%) Store 0% 2.2 (3%) Branch 20% 2.4 (27%) Total CPI.5 Example-2 : 设某程序中 FP 操作占 20%, 其平均 CPI=4.0, FPSQR 操作的比例占 4%, 其 CPI=20.0, 其它指令平均 CPI=.20, 现采用两种方法进行优化 : a. 将 FP 操作的 CPI 减为 2; b. 将 FPSQR 操作的 CPI 减少为 2; 问哪一种方法更优? 若同时采用 a b 方法, 系统的加速比多少? [ 解 ]:CPI a =(-20%-4%) CPI 其他 +(20% CPI FP ) +4%CPI FPSQR = % 2+4% 20 = 2. CPI b =(-20%-4%) CPI 其他 +20% CPI FP +4% CPI FPSQR = %+2 4%=2.3 所以, 采用方法 a 更优 Speedup: CPI ( 原 ) S CPI ( 新 ) ( 20% 4%) CPI ( 20% 4%) CPI 其它 其它 20% CPI 20% CPI 76%.20 20% 4 4% %.20 20% 2 4% 2 FP 4% CPI 4% CPI FPQSR FP FPQSR Principle3: Locality of Reference Thumb: only 20% instructions consume 80% execution time. Programs access a relatively small portion of the address space at any instant of time. Two different types of locality: Temporal Locality (locality in time): If an item is referenced, it will tend to be referenced again soon (loops, reuse, etc.) Spatial Locality (locality in space/location): If an item is referenced, items whose addresses are close by tend to be referenced soon (straight line code, array access, etc.).6 Measuring and Reporting Performance Metrics-: Throughput Execution time is the best measure of computer performance! Two issues: Metrics how do we describe in a numerical way for the performance of a computer? What tools do we use to find those metrics? Application Programming Language Compiler Data BUS Datapath Control Function Units Transistors Wires Pins Operations per second (millions) of Instructions per second: MIPS (millions) of (FP) operations per second: MFLOP/s Megabytes per second Cycles per second (clock rate) 6

7 Methods For Predicting Performance Benchmarks, Traces, Mixes Hardware: Cost, delay, area, power estimation Simulation (many levels) ISA, RT, Gate, Circuit Queuing Theory to predict the load Fundamental Laws /Principles Index : MIPS(Million Instructions Per Second) Features: The speed of a given CPU depends on many factors, such as the type of instructions being executed, the execution order and the presence of branch instructions. Therefore, MIPS only can be useful when comparing performance between processors made from a similar architecture. CPU instruction rates are different from clock frequencies, some instruction may require several clock cycles, or the processor may be capable of executing multiple independent instructions at once. MIPS has not become a measure of instruction execution speed. Effective MIPS speeds are highly dependent on the programming language used. nstructions_per_second Index 2: MFlops(Million Floating point Operation Per Second) Features: )MFlops can only indicate the performance of float point performance, it can t represent overall system performance. 2)MFlops is highly dependent on programming language, instruction sets, e.g *9.3 2 and /9.3 2 are no comparable. Float Point Operations MFlops 6 Execution Time 0 Index 3:Benchmark Test Background: No single numerical measurement can completely describe the performance of a complex device like a computer. The only totally accurate way to measure the performance of your system, however, is to test the software applications you use on your computer system. Benchmark results published by Intel are measured on specific systems using specific hardware and software configurations. Two kinds: component and system. Component benchmarks measure the performance of specific parts of a computer system, such as a microprocessor or hard disk drive, while system benchmarks typically measure the performance of the entire computer system. Multiple Benchmarks for different purposes. E.g. CPU file read/write I/O network and so on. Typical Benchmarks:. Comprehensive: E.g. Dhrystone,Whetstone. 2. Kernel:E.g. Livemore Fortran Kernals, NAS. 3. Matlib:E.g. Linpack, FFT. 4. Application:E.g., SPEC,Perfect, Splash 5. Parallel:NPB, PARKBENCK of NAS You can find LINPACK, LAPACK, BLAS, BLACS, Livemore and other benchmark test tools. SPEC: System Performance Evaluation Cooperative First Round programs yielding a single number ( SPECmarks ) Second Round 992 SPECInt92 (6 integer programs) and SPECfp92 (4 floating point programs) Third Round 995 new set of programs: SPECint95 (8 integer programs) and SPECfp95 (0 floating point) benchmarks useful for 3 years Single flag setting for all programs: SPECint_base95, SPECfp_base95 7

8 Metrics-2: System s Reliability Bathtub Curve. Concepts Error: reasons of fault Fault: bugs, flaws and errors initiate system fault in run-time. Failure: faults cause system disaster, it shows Bathtub curve rule. Fault avoidance: to prevent or hold back fault happening. Fault tolerant: system still can run even error or fault 207 happened. 年 3 月 How to measure Reliability Reliability:probability of no fault, R () t e t failure rate MTTF: Mean Time To Failure MTTF Rtdt e t () dt e t R(t) t 例 : 设系统的平均无故障时间是 0 4 小时, 问该系统正常工作 小时的可靠性是多少? 解 : 因 MTTF=0 4, 故 l=0-4, R()=e =0.990 MTBF = MTTF+MTTR=/ + /u Wherein, MTTR is the average time for recovering failure/u, and u is the recovery rate. MTBF: Mean Time Between Failure Availability of system MTTF / Availability MTTF MTTR / / Summary Website of my courseware: Summary:. Concept: What s computer architecture? 2. Classification of Computer Architecture. 3. Design Principle: Make Common Case fast, Amdahl Law, CPU Performance, Locality. 4. Evaluation Benchmark. Exercises: References: Read the listed Wikipedia s websites. 8