Course overview Computer system structure and operation

Transcription

1 Computer Architecture Week 01 Course overview Computer system structure and operation College of Information Science and Engineering Ritsumeikan University

2 reference information course web site: evaluation mid-term exam 30% final exam 70% books David and Sarah Harris, Digital Design and Computer Architecture also covers digital electronics recommended Hennessy & Patterson, Computer Organization and Design the standard introductory textbook for computer architecture (not as good as Harris, IMO) Hennessy & Patterson, Computer Architecture, a Quantitative Approach similar, but with more mathematics 2

3 course outline 01 Computer system structure and operation computer architecture, components of computers, binary numbers and arithmetic 02 How computation is performed by computers (1) assembly languages and machine languages comparison of common programming models, advantages and disadvantages practice: assembling a program 03, 04, 05, 06 Instruction Set Architectures MIPS instruction set and assembly language OS interface: system calls instructions, operands, addressing modes data representation, arrays, structures comparisons, branches, control structures function calls and the stack 3

4 course outline 07 Converting high-level language to executable program low-level equivalents to high-level programming constructs compiler, assembler, linker, loader 08 Mid-term test 09 How to perform computations in computers (2) CPU architecture, data path, the control system, instruction execution 10, 11 Arithmetic operations in computers ALU: addition, subtraction, multiplication, division floating-point arithmetic 12 Performance evaluation of computer systems response time, throughput, clocks-per-instruction, benchmarks, power efficiency 4

5 course outline 13 Hardware optimisation instruction pipelines, overlapped execution, stalls branch prediction the ARM solution: conditional instructions vs. explicit control structure 14 The Memory Hierarchy hierarchical memory systems locality principle and caching paged memory systems, virtual memory, segmented addressing the MMU 15 I/O, communication, and processes interrupt handling peripheral devices, DMA timers, the process abstraction, preemptive multitasking 5

6 historical perspective ENIAC (Pennsylvania, 1948) Electronic Numerical Integrator And Computer program operations hard-wired together later modified to store program in ROM a little like a modern microcontroller e.g., Arduino program is in flash memory Small-Scale Electronic Machine (Manchester, 1948) program stored in RAM first stored-program computer can be treated as data by other programs basis for all modern general-purpose computers 6

7 historical perspective Manchester Small-Scale Electronic Machine (SSEM) 32-bit words, 32 words of memory, 1 register, 7 instructions 7

8 technological enabler: semiconductors 1940 electromechanical valves (vacuum tubes) need to warm up need high voltages (300V) inefficient as switches generate lots of heat slow unreliable ENIAC: 17,000 valves several failures per day transistors 1955 solid-state instant-on low voltage (5V) fully on, fully off generate little heat fast reliable 8

9 technological enabler: large-scale integration 1965 integrated circuit 1971 microprocessor 1 component = several logic gates 1 component = entire CPU 1971 Busicom 1975 microcomputer several components = calculator several components = computer 9

10 first microprocessor: Intel 4004 (1971) 4-bit words, 640 bytes of memory, 16 registers, 46 instructions 10

11 first microprocessor: Intel 4004 (1971) 2, 300 transistors, 10, 000 nm process, 12 mm 2 die, 0.74 MHz clock 11

12 semiconductor progress Moore s Law 1 predicted in 1965 that, for a single (cost-effective) integrated circuit, the number of transistors doubles every months 1 Gordon Moore, co-founder of Intel and Fairchild Semiconductor 12

13 modern microprocessor: Intel Core i7-6950x (2016) 3, 200, 000, 000 transistors, 14 nm process, 246 mm 2 die, 4, 000 MHz clock 13

14 modern microprocessor: Intel Core i7-6950x (2016) 10 CPUs, 25 MB on-chip (L3 cache) memory, 4 parallel main memory channels 14

15 modern microprocessor: Intel Core i7-6950x (2016) 64-bit words, 128 GB memory, 16 registers 2 22 cores, 3, 683 instructions 15

16 modern microprocessor: Intel Core i7-6950x (2016) approximately 220 dies (CPUs) per 300 mm wafer 16

17 cost original cost equivalent today ENIAC $500,000 $6,300, $5 $26 (manufacture) 4 GHz Intel Core i7-6950x $1570 (retail) 25 MHz MIPS 32-bit microcontroller <$1 current CPU cost depends on yield (how many good CPUs per wafer manufactured) the lower the yield, the higher the price yield decreases with smaller feature (transistor) size smaller features are damaged by smaller defects defect density increases as defect size decreases larger die size larger die more likely to include a defect yield figures are difficult to obtain, but can be inferred from vendor pricing 17

18 components of computers control instruction data address memory program data registers ALU CPU I/O devices ALU MMU cache multiple function units (parallelism) virtual memory, multiprocessing, protection memory latency mitigation (speed) pipeline DMA I/O interrupts overlapped execution (parallelism) background I/O transfer (parallelism) memory-mapped (simplicity) asynchronous I/O, timesharing 18

19 architectural evolution driven (and/or constrained) by several factors operating system (software) progress more sophistication better user interface for programmers memory protection, asynchronous input/output, timesharing technology (semiconductor) progress more transistors more functionality per CPU instruction-level parallelism increased throughput instruction pipeline overlaps execution (more instructions issued per second) multiple ALUs execute multiple computations in parallel until power (heat) becomes a problem, then use transistors for... reduced latency less waiting for data in memory store more program/data in fast memory (cache) inside the CPU decreases access time without (significantly) increasing power 19

20 components of computers control pipeline MMU data address memory program cache data registers DMA ALU CPU I/O devices ALU MMU cache multiple function units (parallelism) virtual memory, multiprocessing, protection memory latency mitigation (speed) pipeline DMA I/O interrupts overlapped execution (parallelism) background I/O transfer (parallelism) memory-mapped (simplicity) asynchronous I/O, timesharing 20

21 technology perspective conventional wisdom (1980) modern reality (2000+) multiply is slow memory is fast power is free transistors are expensive more transistors = more parallelism CPU speed doubles every 1.5 years multiply fast (2 4 cycles) memory slow ( 200 cycles) = memory wall power is expensive (heat) transistors free (cannot use all at once) = power wall diminishing returns (communication delay, linear programs, shared memory) = parallelism wall memory + power + parallelism walls = brick wall further progress depends on other approaches explicit parallelism, vector/matrix maths in GPGPU, distributed algorithms,... 21

22 review: binary numbers unsigned binary numbers a = = A 5 = = signed 2 s complement numbers (2 11 ) a = = s complement negation form 1 s complement (invert each bit), then add 1 22

23 review: binary arithmetic addition just like decimal subtraction just like decimal, but borrow 2 (not 10) when necessary or negate (2 s complement) the subtrahend, then add: x y x + ( y) multiplication just like decimal, sum of partial products division just like decimal, long division 23

24 resources books epl-share: Shared/Books/Microprocessors MIPS Assembly Language Programming (Robert Britton) See MIPS Run (Dominic Sweetman) epl-share: Shared/Books/Computer Architecture Digital Design and Computer Architecture (Harris, Harris) Computer Organization and Design (Hennessy, Patterson) Computer Architecture a Quantitative Approach (Hennessy, Patterson) assembly language practice MIPS machine simulator (assembly language programming) install from Homebrew (recommended): brew install spim with GUI: Omega2+ (a real MIPS-based computer running Linux) docs/omega2p.html I have several of these; if you want one, let me know 24

25 homework next week we will practice running a small machine code program using the MIPS architecture in preparation, please install Spim, a MIPS simulator, on your laptop recommended: using Homebrew: brew install spim this version runs programs from a terminal command line programs behave very much like real, compiled MIPS programs do simple and convenient, if you are comfortable with what the machine is doing or from: download the most recent installer for Mac or Windows this version has a graphical user interface GUI displays the machine registers, program, and memory contents step the program one instruction at a time, see what effect it has good for debugging, if you cannot envision what the machine is doing 25

26 glossary asynchronous an event that occurs with independent timing cache fast memory that stores often-used values, reducing the time needed for their retrieval channels (memory) an address, data and control bus connecting main memory to the CPU. Each channel operates independently, with a limited maximum speed (transfers per second). The more channels a CPU has, the more data can be transferred to/from memory each second. core (processor) the part of the CPU that executes programs. Multiple cores often share a single die. A quad-core processor runs four independent programs at the same time, but all four share the remaining resources of the CPU such as the cache and I/O system. defects physical imperfections that occur in the manufacture of semiconductors. They can be caused randomly by contamination (e.g., by dust or chemicals) of the die, or systematically by optical problems with the photomask or exposure process. die one complete semiconductor device, such as an integrated circuit. Usually produced in bulk on a single wafer that is then cut into individual dies. 26

27 feature the smallest detail present in a semiconductor device. (E.g., for a transistor that contains three elements gate, source and drain, the feature size is typically defined as the width of the channel separating source and drain, or the size of the gate.) function units a unit that performs a specific arithmetic operation. Examples include integer addition/subtraction, integer multiplication, floating point division, etc. Multiple function units within a single CPU can operate in parallel, allowing more than one computation to be performed at the same time. hard-wired a function or program that is described by hardware connections. To modify the function or program, the hardware itself must be physically modified. instruction-level parallelism the computations or instruction execution processes that can be performed in parallel with no explicit indications from the programmer. For example, if two independent integer addition instructions occur next to each other, and two integer addition function units are available and unused, both addition instructions can be issued at the same time and executed in parallel, one in each unit. integrated circuit a semiconductor device that integrates multiple transistors onto a single die to perform a complex logic function. 27

28 issue (instruction) commencing the execution of an instruction. In a pipelined CPU, instruction issue is the process of inserting the next instruction into the first stage of the pipeline. If multiple function units are available, multiple instructions using different function units can be issued at the same time. latency the delay between requesting something and finally receiving it. memory-mapped a technique for referring to things other than memory as if they were memory locations, by associating them with a memory address and then using the address and data bus for communication. The control and data registers associated with I/O devices are typically memory-mapped so that programs can read and write them as if they were normal memory locations. In framebuffer-based graphics cards, the entire framebuffer can be memory-mapped and appear to programs as a block of normal memory, even though its contents are stored inside the graphics card. multiprocessing running multiple programs on a computer such that each has the illusion of owning the entire machine, unaware of the presence of any other programs (unless explicit steps are taken to communicate with those other programs). 28

29 pipeline a linearisation of the basic machine cycle into independent steps that can overlap. A basic machine cycle of five steps (fetch, decode, read operands, execute, write result) can be converted into a five-stage pipeline. One instruction then requires five cycles to execute, but each cycle can be much shorter and five consecutive instructions can be executing in parallel, one in each stage of the pipeline, with a corresponding increase in the instruction issue rate. process (semiconductor) the set of manufacturing techniques designed to create semiconductors with a given feature size. A 14 nm process creates semiconductors with 14 nm features. protection (memory) preventing a program from accessing the memory of another program, or even its own memory in inappropriate ways. solid-state a generic term for the kinds of devices that can be manufactured using semiconductor technology. Solid-state devices have no moving or discrete mechanical parts, unlike relays or valves (vacuum tubes) for example. stored-program a computer that uses normal memory to store and execute its programs. throughput (CPU) the number of instructions that can be executed each second. 29

30 timesharing a computer system that guarantees each program receives a fair share of CPU time. transistor a semiconductor device that can act as a digital switch (or as an analogue amplifier). Many kinds of transistor exist, but almost all are three-terminal devices in which a voltage applied to one terminal controls the amount of current flowing between the other two terminals. (In a bipolar transistor, the voltage on the base controls the current flowing between the collector and the emitter. In a field-effect transistor, the voltage on the gate controls the current flowing between the source and the drain. Modern digital devices typically use field-effect transistors.) virtual memory a technique for decoupling the program s view of memory from the physical arrangement of memory in the computer. Virtual memory allows all programs to use the same range of addresses without conflict, and permits parts of programs to be absent (or temporarily swapped out ) from physical memory which can then be assigned to other programs. wafer a circular slice of mono-crystalline semiconductor material (typically silicon) on which electronic circuits are formed by optical and chemical processes. yield the proportion of devices (dies) on a wafer that have no significant defects and which operate correctly. 30