Cost-Effective Parallel Computing David Wood and Mark Hill University of Wisconsin IEEE Computer 1995

1
Cost-Effective Parallel ComputingDavid Wood and
Mark HillUniversity of WisconsinIEEE Computer
1995

• Presented by Eric Wheeler

2
Theme

• Obtaining sub-linear speedups when implementing a
  multiprocessor can still be worthwhile.

3
Balanced Systems

• Amdahls Dictum
• For each MIPS of processing power need 1 MB
  memory.
• Converse should be true, too.

4
Formalism

• Speedup simply needs to exceed costup to be
  cost-effective

5
Remember Memory

• Previous views are processor centric. Memory can
  dominate cost!

6
Generalization

• When memory dominates, costups can be much
  smaller than linear in the number of processors.

7
Takeaways

• Balanced systems should be a goal.
• Speedups do not have to be linear to be effective
  thanks to the peripheral costs of adding another
  whole machine.

8
Questions

• What should you consider as cost when deciding
  if your project design was effective in light of
  this paper?
• What other things can dominate the cost?
• What happens to this theory if memory becomes
  essentially free (e.g. low power, high density,
  cheap nanoarrays)?