Skip to content

1 why parallelism

🕰️ Historical Context of Parallel Computing

约 165 个字 6 张图片 预计阅读时间 1 分钟 共被读过 次

■ 1970s–2000s: Supercomputers & Databases

  • C.mmp at CMU (1971): 16 PDP-11 processors.
  • Cray XMP (1984): 4 vector processors.
  • Sun Enterprise 10000 (1997): 16 UltraSPARC-II processors.

■ Inflection Point (2004)

  • Power Density Wall: Intel abandons frequency scaling, shifts to multi-core CPUs.

💡 Key Concepts

■ Speedup Formula

\[\text{Speedup}(P) = \frac{\text{Execution Time (1 processor)}}{\text{Execution Time (P processors)}}\]

Demo Observations:
1. Demo 1: Communication limits speedup.
2. Demo 2: Work imbalance reduces efficiency.
3. Demo 3: Communication dominates computation.

🖥️ Modern Parallel Hardware

■ Apple Products

  • Mac Pro: 12-core Intel Xeon E5.
  • iPad Retina: 2 Swift cores.

■ Supercomputers

  • Titan (#2 Supercomputer): 18,688 AMD CPUs + 18,688 NVIDIA GPUs.

🧩 Course Themes

  1. Scaling Parallel Programs:
    - Decomposition, work assignment, communication.
  2. Hardware Efficiency:
    - Performance vs. cost vs. power.
  3. Post-2004 Shift:
    - Maximize performance per Watt instead of raw speed.

🚨 Key Takeaways

  • Single-thread performance growth is stagnant → Parallelism is essential.
  • Writing parallel code is challenging but unlocks immense computational power.
  • Efficiency matters: 2x speedup on 1010 processors is not impressive.

📢 Welcome to 15-418!