From ce5893dc8b77bae37a70a6b584057d29365941dc Mon Sep 17 00:00:00 2001 From: Alex Razoumov Date: Mon, 16 Sep 2024 10:49:23 -0700 Subject: [PATCH] make the keypoints in 14-parallel-case-study.md more specific --- episodes/14-parallel-case-study.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/episodes/14-parallel-case-study.md b/episodes/14-parallel-case-study.md index 75e23f2..a5e4f6c 100644 --- a/episodes/14-parallel-case-study.md +++ b/episodes/14-parallel-case-study.md @@ -342,5 +342,9 @@ coforall (i,j) in {1..n,1..n} by (stride,stride) { // 5x5 decomposition into 20x We will study data parallelism in more detail in the next section. ::::::::::::::::::::::::::::::::::::: keypoints -- "There are many ways to implement task parallelism for the diffusion solver." +- "To parallelize the diffusion solver with tasks, you divide the 2D domain into blocks and assign each block + to a task." +- "To get the maximum performance, you need to launch the parallel tasks only once, and run the temporal loop + of the simulation with the same set of tasks, resuming the main task only to print the final results." +- "Parallelizing with tasks is more laborious than parallelizing with data (covered in the next section)." ::::::::::::::::::::::::::::::::::::::::::::::::