Carnegie Mellon Parralel Computing Notes on Lecture 6

How much to steal how to detect program termination

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: mbalance! Time P1 P2 P3 P4 Done! One possible solution to imbalance problem: Divide work into a larger number of smaller tasks - Hopefully “long pole” gets shorter relative to overall execution time - May increase synchronization overhead - May not be possible (perhaps long task is fundamentally sequential) CMU 15-418, Spring 2014 Smarter task scheduling Schedule long task rst to reduce “slop” at end of computation Time P1 P2 P3 P4 Done! Another solution: smarter scheduling Schedule long tasks rst - Thread performing long task performs fewer overall tasks, but approximately the same amount of work as the other threads. - Requires some knowledge of workload (some predictability of cost) CMU 15-418, Spring 2014 Decreasing synchronization overhead ▪ Distributed work queues - Replicate data to remove synchronization Subproblems (a.k.a. “tasks”, “work to do”) Set of work queues (In general, one per worker thread) Steal! Worker threads: Pull data from OWN work queue Push new work to OWN work to queue When localwork queue is empty... STEAL work from another work queue T1 T2 T3 T4 CMU 15-418, Spring 2014 Distributed work queues ▪ Costly synchron...
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