Previous | Next --- Slide 5 of 43
Back to Lecture Thumbnails
Metalbird

A speedup should be measured against the performance of the best sequential program. When designing an algorithm specifically for multiple cores, performance may be poor for one core and spectacular for multi-core showing a huge speedup. However, this is irrelevant because no one should use the multi-core version on a single core if they were trying to get the best performance possible. So usually the most meaningful speedup would be comparing the best sequential program to a parallel version on multiple cores.

zale

There was a paper submitted at USENIX to highlight that idea. They also touch on the fact that sometimes optimizations will make a program look less scalable, but be actually faster on any core count, just by reaching better performance on low core counts.