Fundamentals of Quantitative Design and Analysis


Introduction

  • Two significant changes in computer market place
    • Virtual elimination of assembly language programming
    • The creation of standardized, vendor-independent operating systems
      • UNIX, Linux, …
  • New set of architectures with simpler instructions
    • RISC (Reduced instruction set computer) architecture
    • Two critical performance technique:
      • Instruction level parallelism
      • Use of caches
    • Challenges to 80x86 instructions
    • ARM, becoming dominant
  • Fourfold effect of dramatic growth of computer market place
    1. Significantly enhanced capability of personal users
    2. New classes of computers
      • Personal computers
      • Workstations
      • Smart cell phones
      • Tablet computers
      • Warehouse computers
      • Supercomputers
    3. Moore’s law-driven hardware renaissance
    4. Software development
      1. Performance-oriented languages: C, C++
      2. Managed programming languages: Java, Scala
      3. Scripting languages: JavaScript, Python
      4. Programming frameworks: AngularJS, Django
      5. Interpreters with just-in-time compilers
      6. Trace-based compiling
      7. Software as a Service: SaaS
      8. Internet
      9. Application: Speech, sound, images, videos
        • Google translate running on warehouse-scale computer (WSC)
  • End of hardware renaissance
    • Dennard Scaling: constant power density for smaller transistor dimensions
    • Moore’s Law: the number of transistors on a microchip doubles every two years.
    • Started to use multiple cores
      • Instruction level parallelism:
        • Compiler and hardware conspire to exploit ILP.
        • No engagement of programmers
      • Data level parallelism
      • Thread level parallelism
      • Request level parallelism for WSC
    • Amdahl’s Law
      • prescribes practical limits to the #cores per chip
    • Thus, ”The only path left to improve energy-performance-cost is specialization”
  • Growth in processor performance since the late 1970s

Classes of Computers

Internet of Things (IoT), Embedded Computers

  • 8/32-bit for low cost devices (Microwaves, Washing machines)
  • 64-bit for high-end product (Cars, Network switches)

Personal Mobile Devices (PMD)

  • Responsiveness and predictability
  • Real-time performance
  • Minimize memory and energy consumption

Desktop Computing

  • Benchmarking + Web-centric, interactive apps

Servers

  • Availability: Open 7 days, 24 hours
  • Scalability: Scale up computing capacity, the memory, the storage, the I/O bandwidth
  • Efficient throughput (Overall Perf.) > Responsiveness (Individual Perf.)

Clusters/Warehouse-Scale Computers

  • The growth of Software-as-a-Service (SaaS)
  • Clusters:
    • Collection of desktop computers or servers connected by local area networks to act as a single larger computer
    • Each node run its own operating system, and nodes communicates using a networking protocol
  • WSCs: The largest of the clusters
    • Use inexpensive, redundant components compared to the servers
  • Price-performance and power
  • Availability: Peak hours for Christmas!
  • Supercomputer
    • Expensive, floating point performance,
    • Running large, communication-intensive batch programs

Class of Parallelism and Parallel Architectures

  • Parallelism in applications:
    • Data-level parallelism
    • Task-level parallelism
  • Parallelism in computer hardware
    • Instruction-level parallelism
      • Pipelining, Speculative execution
    • Vector architectures, graphic processing units (GPUs), multimedia instruction sets
      • Data-level parallelism by applying a single instruction to a collection of data in parallel
    • Thread-level parallelism
      • DLP & TLP in a tightly coupled hardware model with interaction hardwares
    • Request-level parallelism
      • Parallelism between largely decoupled tasks
  • Flynn’s (1966) taxonomy
    • SISD: ILPs such as superscalar and speculative execution
    • SIMD: Vector architectures, graphic processing units (GPUs), multimedia instruction sets
    • MISD: eg. Systolic array
    • MIMD: DLP & TLP & RLP, Tightly/loosely coupled multicore

Defining Computer Architecture

Instruction Set Architecture: The Myopic View of Computer Architecture

  • ISA
    • The actual programmer-visible instruction set
    • A boundary between the software and hardware
    • 80x86, ARMv8, RISC-V
    • RISC-V
      • A large set of registers
      • Easy-to-pipeline instructions
      • A lean set of operations
  1. Class of ISA
    • General-purpose register architectures: Nearly all ISAs
      • Register-memory ISAs (80x86)
      • Load-store ISAs (ARMv8 & RISC-V)
  2. Memory addressing
    • Mostly, byte addressing
    • Some, objects should be aligned
  3. Addressing Mode
    • Addressing modes specify the address of a memory object
    • eg. In RISC-V, Registers / Immediate / Displacement
  4. Types and Sizes of operands
    • ASCII, Unicode, INT, word, FP32, FP64 ….
  5. Operations
    • Data transfer, arithmetic logical, control, floating point
  6. Control flow instructions
    • Conditional branches, unconditional branches, jumps, procedure calls, and returns
  7. Encoding an ISA
    • Fixed length vs Variable length
  • RISC-V registers, names, usage, and calling conventions

Genuine Computer Architecture: Designing the Organization and Hardware to Meet Goals and Functional Requirements

  • Implementation = Organization (Microarchiecture)+ Hardware
  • Architecture = ISA + Organization (Microarchiecture) + Hardware

  • More information in Appendix.A (HP) Instruction Set Principles)

  • Technology: IC logic technology, DRAM, Flash, Disk, Network
  • Performance
  • Scaling of transistors and wires
  • Power and energy
  • The shift in computer architecture because of limits of energy:
    • Dark silicon
    • Domain-specific processors
  • Cost
    • Time, volume, commoditization
    • Cost of manufacturing vs operation


Dependability

  • Is a system operating properly?
  • How a failure in a specific level of computer influence other levels, esp. for application level?
  • Service level agreements (SLAs) or Service level objectives (SLOs)
  • Module reliability
    • Mean time to failure (MTTF)
    • Mean time to restore (MTTR)
  • Module availability = MTTF/(MTTF+MTTR)

Measuring, Reporting, and Summarizing

  • Benchmarks
    • Kernels: small, key pieces of real applications
    • Toy programs
    • Synthetic benchmarks
    • Benchmark suites: Collections of benchmarks
      • SPEC (Standard Performance Evaluation Corporation)
  • Processor performance
    • CPI: Clock cycles per instruction
    • IPC: Instructions per clock

Quantitative Principles of Computer Design

  • Take Advantage of Parallelism
  • Principle of Locality
  • Focus on Common Case
  • Amdahl’s Law

Reference

Notes Mentioning This Note

Table of Contents


Share on: