Rerun PR visualizations with iterations info
crozhon opened this issue · comments
Finished a lot sooner than I expected, so I added the results to io/gunrock-output/pr-iteration-rerun/. They have the tag
pr-iteration-rerunand include the new fieldelapsed-iterations.
elapsed-iterationsis anum-runssized array. If it exists, it specifies the actual number of iterations performed for every run of the problem. On the other hand, the fieldmax-iteris an upper bound on the number of iterations performed and doesn't specify how many were executed.
Do we have any other programs, besides Pagerank, that need the number of iterations that were executed for a single run?
Originally posted by @crozhon in gunrock/gunrock#813 (comment)
Could you try redoing PR with this new elapsed-iterations field?
@crozhon would like another pair of eyes on this
compare https://gunrock.github.io/docs/#/analysis/results_pr?id=pr-throughput-mteps-for-push-vs-pull-separating-out-options to what's attached
It looks like you only reran K40 with pull and the numbers are considerably worse, which is expected since we are reporting per-run numbers and it looks like we now have 7 iterations not 50.
If that's the case, can you do a full run (V100 and K40 both, push and pull both) and delete all the inaccurate runs in v1-0-0/pr and replace them with the good runs? You can delete pr-iteration-rerun as well. Everything is in source control so we don't lose any info, but the bad files shouldn't be in a current clone of the outputs repo.
gunrock_primitives_pr_push_pull_mteps.pdf
That's right. The iterations are often considerably smaller than 50, so it makes sense. I only reran with K40 and pull originally, but I started up a full run two nights ago, and it's finishing up today.
Sounds good. I'll push them and delete the old ones as soon as it's ready.
Do you know why K40C doesn't have any runs with pull=false in the directory? See https://github.com/gunrock/io/tree/master/gunrock-output/v1-0-0/pr/K40c I'm getting some really long process-times for K40c and the other graphs.
Did you originally compare Titan V and V100 instead?
Do you know why K40C doesn't have any runs with pull=false in the directory? See https://github.com/gunrock/io/tree/master/gunrock-output/v1-0-0/pr/K40c I'm getting some really long process-times for K40c and the other graphs.
Did you originally compare Titan V and V100 instead?
I don't have an answer for this. @neoblizz ?
Do you know why K40C doesn't have any runs with pull=false in the directory? See https://github.com/gunrock/io/tree/master/gunrock-output/v1-0-0/pr/K40c I'm getting some really long process-times for K40c and the other graphs.
Did you originally compare Titan V and V100 instead?I don't have an answer for this. @neoblizz ?
There was a bug for PR on SM_35 (K40), which is why I never ran a push or pull-based run on K40. The recent pull only results are only there because I ran those for Carl's paper. So, we really need a complete run of push-pull PR on K40.
(Bug may have been gunrock/gunrock#757 or gunrock/gunrock#682)
That's correct. There were still issues running out of memory with push.