CS251 - Computer Organization and Design - Spring 2008

Lecture 2 - Performance

Practical Details

1. Course web pages
2. Excessive collaboration
   • Copying
   • Use of external material without acknowledgement
     • books
     • web pages
     • newsgroups

     If in doubt state what you did

   Good rule of thumb: talk together but write independently

   Penalties are severe

3. Read chapter 1
   • Do exercises 1.1 to 1.45 (estimated time 15 minutes)

     These are things you are expected already to know

   • Look through exercises 1.46 to 1.54

     These are things it will be very useful for you to know

4. Read sections 4.1 to 4.3; skim sections 4.4 to 4.6
   • Exercises 4.1 to 4.18, 4.45 and 4.46 could be assignment or exam questions.

Performance of Computer Systems

In the box on page 271, the authors write, `Execution time is the only valid ... measure of performance.'

This is not true!!

1. real-time systems
2. low power systems
3. portable systems
4. robust systems
5. systems with user interfaces
6. not to mention `entertaining systems' or `beautiful systems' or `systems that make you feel cool'.

Exercise for the reader. Write an essay on the social distortions that make it possible for the authors to write what they did, and presumably, believe, it.

So, how do we measure performance?

1. Define an objective function.
   • Might be as simple as `execution speed' x `battery life'
   • Might be as complex as transactions per hour using our mix of transactions
2. Measure the objective function in different system configurations

Execution time is commonly used as a proxy for other measures.

• The drunk under the lamppost.
• Goodhart's law.

What affects execution time?

Features of the computer

1. clock speed
2. instruction set
3. compiler efficiency
4. memory speed
5. I/O bandwidth

Features of the workload

1. instruction mix
2. computation versus memory versus I/O
3. data locality

Possible computer solutions

• (1&2) two design choices
  1. CISC: complex instruction set, many instructions, fast clock, multiple cycle execution times
  2. RISC: reduced instruction set, few simple instructions, fast clock, single cycle execution
• (3) two design choices
  1. Procedural: simple to compile, not much to optimize
  2. Object-oriented: hard to compile, heavy optimization needed
• (4) multi-level cache
• (5) co-processors

Benchmarks

Collections of code to be executed, standard sets of problems to solve.

• Appear to be objective
• Can be officially made into standards
• Design for the benchmark: Goodhart's law

Return to:

• Bill Cowan's lecture notes for CS251 in W08
• Bill Cowan's Winter 2008 CS251 page
• Bill Cowan's CS251 page
• Bill Cowan's teaching page
• Bill Cowan's home page