1. 병렬처리시스템
2005년도 2학기
채 수 환

2. Textbook & Evaluation Policy
Text : Pipelined and Parallel Computer Architectures by Sajjan G. Shiva, Harper Collins 1996.
A Reference Book : Introduction to Parallel Computing by Ananth Grama, et al. , Pearson ,2003
Evaluation Policy: Mid-term Exam. 40% Final Exam. 40% Report %
Attendance %

3. How to Prepare Your Future
Do not hurry up for success in your life.
Make your plan for your future.
Spare your time.
Be an expert in your field.
Have some meditation time each day.

4. Chapter 0 Introduction
Performance and cost are two major parameters for evaluation of architectures.
The waves of computing
The 1st wave in the 1960s was dominated by mainframe.
The 2nd wave in the 1970s was belonged to the minicomputer systems.
The 3rd wave in the 1980s was that of microcomputers.
These three waves are considered evolutionary approaches.
Parallel processing is considered the 4th wave of computing. It is considered revolutionary approach.

5. I.1 Computing Paradigms
Three paradigms of computing structures (Figure I.1)
serial
pipelined
parallel

6. Three Paradigms

7. Example I.1
T: the time to complete a subtask, N: the number of tasks, A task consists of 4 subtasks.
The time achieved by the serial paradigm.
Serial: 4NT
The speedup achieved by the pipelined paradigm
4NT / {4T +(N-1)T}
The speedup achieved by the parallel architecture
4NT / P, where P is the number of processors

8. I.1 Computing Paradigms (continued)
Trivial parallelism: A job can be partitioned into independent tasks.
In practice, clean partitioning is not possible.
A considerable communication and task coordination overhead is needed in parallel processing.

9. I.2 The need for pipelined and parallel processing
Figure I.2 shows the summary of the fields requiring high performance computation.

10. Grand Challenges

11. I.2 The need for pipelined and parallel processing(continued)
The overall processing speed depends on two factors:
the computing speed of the processors
the communication speed between their interconnection structure
The communication overhead is the function of the bandwidth and the latency of the interconnection structure.
Because of the low cost of microprocessors, it is possible to build low-cost high-performance systems utilizing a large number of them.

12. I.2 The need for pipelined and parallel processing(continued)
MPP(Massively Parallel Processor) architectures consists of a large number of processors interconnected by high bandwidth and low-latency network.
The major issues of pipelined and parallel computing architectures:
hardware: scale-up
algorithm: efficient parallel algorithms
language: programming languages for applying parallel algorithms
compilers and other programming tools: compilers to translate parallel programming language into optimized object code, simulators, debuggers
operating systems: O.S to make control of a parallel processing architecture efficiently.
Performance evaluation: methods to evaluate speedup based on various metrics for parallel processing systems.

13. I.3 Overview Part 1: Preliminaries (Chapter 1 and 2)
Uniprocessor architecture overview
Styles of architecture
Part 2: Pipelined Architectures (Chapter 3 and 4)
Pipelining
Vector processors
Part 3: Parallel Architectures (Chapter 5 and 6)
Array processors
Multiprocessor systems
Part 4: Experimental Architectures (Chapter 7 and 8)
Dataflow architecture
Distributed processing systems, the 5th or 6th generation computers