[Amdahl was an optimist]

In practice, applying an optimization to one part

of a codebase can require overheads that did not exist in the original

execution or even have negative impacts on logically unrelated portions of

the code. Consider an

already parallel

code whose execution consists of

SERIAL (S), PARALLEL-COMPUTE (PC), and PARALLEL-SYNC (PS) regions,

where, with 1 thread, S=20%, PC=75%, and PS=5% of execution time.

Assume that the amount of execution required for PS grows linearly with the

number of threads, but is itself 80% parallelizable, and that PC receives a 1:1

linear speedup with the number of threads. Find

1) the number of threads that maximizes the performance of this code, and

2) the speedup of a 100% parallelizable PS over 1) for the same number of

threads that maximizes the performance of 1).

(note no knowledge of parallel threading or synchronization is required to

solve this problem it is a direct application of Amdahls law)