Carnegie Mellon Parralel Computing Notes on Lecture 6

05 in amdahls law eqn cmu 15 418 spring 2014 static

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: tra work performed to increase parallelism, manage assignment, etc. ▪ We are going to talk about a rich space of techniques - TIP #1: Always do the simplest thing rst, then measure/analyze “It scales” = your code scales as much as you need it to (if you anticipate only running low core count machines, it may be unnecessary to implement a complex approach that created hundreds or thousands of pieces of independent work) CMU 15-418, Spring 2014 Balancing the workload Ideally all processors are computing all the time during program execution (they are computing simultaneously, and they nish their portion of the work at the same time) Time P1 P2 P3 P4 Recall Amdahl’s Law: Only small amount of load imbalance can signi cantly bound maximum speedup P4 does 20% more work → P4 takes 20% longer to complete → 20% of parallel program runtime is essentially serial execution (clari cation: work in serialized section here is about 5% of a sequential program’s execution time: S=.05 in Amdahl’s law eqn) CMU 15-...
View Full Document

This document was uploaded on 03/19/2014 for the course CMP 15-418 at Carnegie Mellon.

Ask a homework question - tutors are online