Unformatted text preview: ess signiﬁcant
• In this class, we’ll mainly be concerned with primary
effects (asymptotic analysis)!
• In the real world, secondary effects are also often
worth paying attention to (after the primary ones)!
UW, Autumn 1999! CSE 373 – Data Structures and Algorithms! Brad Chamberlain! “Returning to the present...” In Parallel Compu3ng, Constants Maher • AsymptoEc analysis (big- O notaEon) is crucial in parallel compuEng, as in tradiEonal compuEng • However, constant factors also maber – in parEcular, we’ll be running on a constant number of processors • anecdote from computaEonal chemist colleague – also, since performance is a primary moEvator for parallel compuEng, we typically want to squeeze out as many overheads as possible CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 53 Measuring Parallel Computa3ons (Directly) Timings: How long did the program take to run? – typically measured in wallclock seconds (or fracEons thereof) Performance: At what rate is the program running? – e.g., FLOPS (ﬂoaEng point operaEons per second) • or simply OPS – or something more domain- speciﬁc: • graph codes: TEPS (traversed edges per second) • memory bandwidth: GB/s (gigabytes per second) • table updates: GUPS (giga- updates per second) CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 54 Measuring Parallel Computa3ons (Rela3vely) Speedup: How does the parallel execuEon compare to a serial execuEon? Speedupp = Tserial / Tp Linear/Ideal Speedup: Speedupp = p CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 55 Sample Speedup Graph CSEP 524: Parallel ComputaEon Winter 2013: Chamberlain 56 Compu3ng Speedup: The Baseline A key issue: What to use for the serial Eming? – some opEons: • the parallel code running using 1 task/processor? • a serial implementaEon of the same algorithm? • the best serial implementaEon available? – The last is the most ideal/valuable • e.g., if...
View Full Document
This document was uploaded on 04/04/2014.
- Winter '09