目录

Fundamentals

Folien aca 01 / Chapter 1 / Video 1.1.1 - 1.1.7

重点是Performance 和 量化原理，其他是简单的定义。

Trends

Classes of Computers

• Desktop

• Server

• Embedded Computer

  distinguish 2 more classes:

• Personal Mobile Devices

• Warehouse-scale Computers

Computer Architect/Design

Determine the important attributes

Design a machine to

• maximize performance and energy efficiency
• stay within cost, power and availability constraints

Including:

• Instruction set design
• Functional organization (u-architecture)
• Logic design and implementation
• Assist marketing, sales, customers
• Validation and debug support

Computer Architecture

is the science and art of selecting and interconnecting hardware components to create computers that meet functional, performance and cost goals.

• Instruction set architecture
• Organization (micro architecture)
• Hardware (Detailed logic design and packaging tech.)

Trends in Power in IC

Issues:

• Get In: Bring Power to the Chip
• Get out: Remove Heat

Why: Battery Life, Power Consumption, Environment, Power Density

Performance

• Time:

  1. $Performance = \frac{1}{Execution Time}$

     1. Increasing performance means decreasing execution time

        Execution Time can be defined in different ways:

        • Wall-clock Time = response time = elapsed time

          The latency to complete a task. Includes Everything: disk accesses, I/O, OS overhead, …

        • CPU Time *(when talking about the time of a program)

          CPU computing time, not time waiting for I/O or other programs

          • User CPU Time : CPU time spent in the program
          • System CPU Time: CPU time spent in OS

  2. Throughput / Rate = n jobs / second

  multiprogrammed CPU works on other process when current process stalled for I/O. It does not improve the execution time of current process, but improve the throughput.

• Availability: the percentage of time, a system is up and running

• Scalability: the ability of a system to expand performance

• Performance per Watt: the number of flops per watt of power as seen before

Benchmarks

a program used to evaluate performance.

Five types (in order of accuracy high to low):

• Real applications
  • portability problems
  • important activity has to be eliminated
• Modified (or scripted) applications
  • enhance portability 增强可移植性 以便多平台运行
  • forcus on particular aspect of system performance
    • e.g. CUP-oriented benchmarks will remove I/O
  • scripts to reproduce interactive behaviour
• Kernels
  • small, key pieces from real programs (key = consume lots of Ex Time)
  • best used to isolate the performance of certain features
• Toy benchmarks
  • small, easy but not useful ti evaluate perf
• Synthetic benchmarks
  • typical instruction mixes
  • whetstone and dhrystone

Benchmark Suites

a collection of benchmarks that measure the performance with a variety of apps.

Most successful attempt tp create a standard: SPEC (focuses on CPU)

• Processor-intensive Benchmarks:
  • SPEC SPU 2006
• Graphics-intensive Benchmarks
  • SPECviewperf: OpenGL 3D rendering
  • SPECapc: extensive use of graphics
• Server Benchmarks with significant I/O
  • SPECSFS: serve
  • SPECWeb: web

Embedded Benchmarks

Less developed than desktop (CPU+GPU) and server benchmarks. (嵌入式系统各不相同，各自设计各自的benchmark)

One attempt: EDN Embedded Microprocessor Benchmark Consortium (EEMBC)

Kernels and benchmarks that fall into 5 classes

• Automotive/industrial
• Consumer
• Networking
• Office automation
• Telecommunications

Summarizing Performance

• total execution time: summary from all programs in one Computer
• arithmetic mean: $\frac{1}{n}\sum_{i=1}^n{Time}_i$

if the programs are not run equally?

• weighted execution time: $\sum_{i=1}^n{Weight}_i \cdot {Time}_i$

• normalized execution time (Used by SPEC):

  Normalize execution times to a reference machine and take average

  SPECRatio = $\frac{Execution Time_{reference}}{Execution Time}$

  don’t use the arithmetic mean to average normalized execution time

  lead to different results depending on the reference machine

  use geometric mean

Quantitatice Priciples

CPU Performance Equation

Cycle time : reciprocal of CPU frequency/Clock rate f Cycle time = rate$^{-1}$ (e.g. 1 GHz: 1 ns)

IC: Instruction count (or instruction path length) = number of instructions executed

CPI : cycles per instruction = Cycle count / IC

so CPU time = Cycle count for a program × Cycle time = IC × CPI × Cycle time

That means, CPU performance equally depends on IC, CPI and frequency.

10% improvement in any of them leads to 10% overall improvement.

We can improve that in:

By designing CPU, it is useful to calculate the number of total CPU clock cycles as:

CPU clock cycles (Cycle count) = $\sum_{i=1}^{n}IC_i \times CPI_i$ (指令数 × 每条指令需要的周期数)

Sometimes we know how many cycles each instruction takes and the frequency of each instruction. The average CPI can be calculated as:

CPI = $\sum_{i=1}^{n}CPI_i\times Freq_{i}$

where
- $CPI_i$ is the average CPI for instruction i
- $Freq_{i} = IC_i / IC$ is the frequency of instruction i

即 知道每个Op的周期，知道每个Op的占比，但不知道每个Op有多少IC

Amdahls’ Law

to calculate the improvement of performance by using a enhancement E

speedup = $\frac{ExTime \quad without \quad E}{ExTime \quad with \quad E}$

if the enhancement E accelerates a fraction F of the whole task by a factor S. The new execution time can be calculated as:

$T_{new} = (1-F) \cdot T_{old} + F \cdot T_{old} / S$

The overall speedup is:

$Speedup = \frac{T_{old}}{T_{new}} = (1-F+\frac{F}{S})^{-1}$

最大speedup (即S趋向正无穷) 是 1-f 的倒数

Implications:

Focus on the Common Case (优化大的部分很赚 80%快2倍 还是 20%快10倍)

Law of diminishing returns: optimizations will have less and less effects

Design Principles

to Make Processors Faster

• Take advantage of parallelism：next chapter
• Principle of locality:
  • Programs spend 90% of their execution time in 10% of their code
  • can predict with reasonable accuracy which instructions and data program will use in near future
• Make the common case fast

Sample Questions

1. The trend of the increment of single-core performance, Causes, Follow-up situation of semiconductor:

   Trend: The 17-year hardware renaissance is over. Single-processor performance improvement has dropped to less 22% per year since 2003.

   Causes:

   6th Edition: The fundamental reason is that two characteristics of semiconductor processes that were true for decades no longer hold. 1. Power density - Dennard scaling ended around 2004 - led to using multiple efficient procesors or cores. 2. - ` the number of transistor ` - Moores' law ended - Q3..

   The combination of

   ■ transistors no longer getting much better because of the slowing of Moore’s Law and the end of Dinnard scaling,

   ■ the unchanging power budgets for microprocessors,

   ■ the replacement of the single power-hungry processor with several energyefficient processors, and

   ■ the limits to multiprocessing to achieve Amdahl’s Law

   caused improvements in processor performance to slow down.

   5th Edition: due to the twin hurdles of maximum power dissipation of aircooled chips and the lack of more instruction level parallelism to exploit efficiently.

   Follow-up: Intel canceled its high-performance uniprocessor projects and joined others in declaring that the road to higher performance would be via multiple processors per chip rather than via faster uniprocessors.

2. three main classes of computers:

   Desktop Computing (largest market), Serves, Embedded computer.

3. Moore’s Law:

   Number of transistors that can be placed inexpensively on an integrated circuit doubles approximately every two years.

4. Benchmark:

   a program used to evaluate performance.

   Real Applications are the best choice of benchmarks.

   SPEC: The Standard Performance Evaluation Corporation (SPEC) is one of the most successful attempts to create standardized benchmark application suites. There are now SPEC benchmarks to cover many application classes.

5. Normalized Execution Time:

   why not arithmetic mean?

   it can lead to different results depending on the reference machine.

   eg.

   	A	B		A	B		A	B
   Program P1	1	10		1.00	10.00		0.10	1.00
   Program P2	1000	100		1.00	0.10		10.00	1.00
   			Arithmetic	1.00	5.05		5.05	1.00
   			Geometric	1.00	1.00		1.00	1.00

6. improve 75% of our application by a factor of 3 xor 50% by a factor of 10.

   $Speedup_1 = ((1-0.75) + 0.75 / 3 ) ^ {-1} = 2$

   $Speedup_2 = ((1-0.5) + 0.5 / 10 ) ^ {-1} = 1.818$

7. improve part of our application by a factor of 4 and want to achieve an overall speedup of 2. How large should the fraction be.

   $ Speedup = 2 = ((1-F) + F / 4 ) ^ {-1} $

   $F = 2/3$

8. In an application 50% instructions take 1 cycle, 20% take 2 cycles, 30% take 2.4 cycles on average. What is the average CPI? This application executes 1.6x10$^9$ instructions and the clock frequency is 3GHz. How much time does the program take?

   $CPI = 50\% \times 1 + 20\% \times 2 + 30\% \times 2.4 = 1.62$

   $IC = 1.6 \times 10^9 \qquad f =3 \times 10^9 Hz$

   $T = IC \times CPI \times f^{-1} = 1.6 \times 10^9 \times 1.62 \times (3 \times 10^9)^{-1} = 0.864 s$

Embedded Systems (Intro.)

Folien none / Appendix E.1 / Video 1.1.8

Embedded Systems

computer system with dedicated function lodged in other devices (the boundary is hard to define.)

Characteristics and Requirements

1. Execute 1 or few applications (designer optimize the system for that app or those apps)
   • lower compatibility battiers (the software is sold together with the hardware)
2. Real-time constraints (complete their tasks on time otherwise iy is considered a failure)
3. Minimize cost
4. Minimize power consumption (powered by batteries and no fans for cooling)
5. Minimize memory usage (the memory can be a substantial portion of the system cost)
6. Physical dimensions (fit in a small device)
7. Interface with physical word

Real-Time Constraints

A real-time performance requirement is one where a segment of the application has an absolute maximum execution time that is allowed.

• Hard real-time system
  • all deadlines are hard (the worst case execution time need to be shorter that the deadline)
  • missed deadline = total system failure
  • e.g. Airbag
• Soft real-time system
  • some deadlines are soft (only the average execution time need to be shorter that the deadline)
  • soft deadline may occasionally be missed
  • e.g. Video Player

Implementation Alternatives

in order of performance power low to high, and also flexibility high to low

1. General purpose processors (GPPs)
   • e.g. ARM processor
   • flexible
   • may not meet the performance requirements of the applications
2. Application-specific instruction set processors (ASIPs)
   • a processor whose instruction set has been tailored to benefit a specific app
   • commercial tools to design an ASIP: Synopsys ASIP Designer, Codasip codasip studio
3. Reconfigurable hardwares - Field programmable gate arrays (FPGAs)
   • can exploit all the parallelism in an app and therefore achieve higher performance
   • can take a very long time
4. Application-specific integrated circuits (ASICs)
   • IC customized for a single particular use
   • takes a long time and a high cost
   • only cost-effective if many devices will be sold

• Previous
  唤醒 矩阵运算
• Next
  机器人学中的状态估计-02