Year: 1989
Type: conference-paper
Abstract: When the number of processors P is less than the number of tasks N in a parallel loop, the loop has to be executed in ⌈N/P⌉ rounds and the last round executes only (N mod P) tasks. In many cases, in the last round all but a few processors are idle, which causes a significant drop in performance. This performance drop becomes more and more detrimental as the number of processors increases. Loop spreading is a technique for restructuring parallel loops so as to balance parallel tasks on multiple processors. A spread loop runs at least as fast as the non-spread loop even when N mod P = 0, and shows no performance drop when N changes. We show how the method keeps the performance of the matrix multiplication and a simplex algorithm from decreasing as the size of input changes.
DOI: 10.1145/76263.76338
Authors: Y. Wu, T. Lewis
Institutions: Nike (United States), Neuromatch, Oregon State University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W2144538029