6.2 Performance metrics for parallel processing
Traditionally, to assess the performance of parallel
implementations, the following metrics are defined:
• Computational cost and
• Speedup.
Computational cost C is defined as the processing
time (in seconds) that a given algorithm consumes.
Then computational throughput T is defined as the
inverse of the computational cost:
T = 1/C
Speedup S measures the reached execution time
improvement and express how fast the parallel
implementation is, compared with the
implementation of reference:
S= T
targ / Tref
where T
targ is the throughput of parallel
implementation under study, and Tref is the
throughput of the sequential implementation in our
case.