We have devised similarity matrices similar to Brad Calder's phase
work based on our component power data, which we name Power
Vectors, and demonstrated phase information with Matrix
Plots. Then we developed a similarity metric based on the original
Manhattan Distance (L1 Norm) idea, which we found more suitable
for power behavior similarity. We used a thresholding algorithm
to group application's execution points into phases and demonstrated
the groups with Grouping Matrices. Afterwards, we have
seen that we can represent the overall program power behavior with
a small set of signature vectors.
We do this research following the hope to be able to use the signatures
for program identification, phase prediction, etc. Also we want
to see how well simulations based on only representative points
can approximate overall power behavior.