Createjoin example int resultloc void addonevoid arg

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: the parallel version is 100x slower on 1 processor, compuEng speedup relaEve to itself is not very helpful –  However, depending on what’s being studied, any of these choices may be reasonable Moral: always pay abenEon to what speedup is computed against! CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 57 Compu3ng Speedup rela3ve to different baselines relaEve to a serial implementaEon CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain relaEve to itself 58 Some More Speedup Graphs These curves are all computed using the idenEcal serial/baseline value CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 59 Measuring Parallel Computa3ons Rela3vely (2) Efficiency: How does the parallel execuEon compare to a linear speedup? Efficiencyp = Speedupp / p Ideal Efficiency: 1.0 for any value of p Arguably the best way to display parallel performance –  it makes the best use of a graph’s area •  doesn’t compress small scales to the lower corner –  yet it’s not used in pracEce as much as you’d think •  in part because it’s hard to achieve 1.0, so it tends to look worse •  also doesn’t give that posiEve sense of “things are trending upward” CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 60 Efficiency Graph (wouldn’t a picture here be great? To the whiteboard!) CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 61 Strong vs. Weak Scalability Strong scaling: Uses a fixed problem size as the number of processors increases. Weak scaling: Grows the problem size with the number of processors Both approaches have their place depending on what you c...
View Full Document

This document was uploaded on 04/04/2014.

Ask a homework question - tutors are online