-
Notifications
You must be signed in to change notification settings - Fork 232
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Qualification tool hook up final output based on per exec analysis #5550
Conversation
Signed-off-by: Ahmed Hussein (amahussein) <a@ahussein.me>
Signed-off-by: Ahmed Hussein (amahussein) <a@ahussein.me>
Signed-off-by: Thomas Graves <tgraves@apache.org>
build |
tools/src/main/scala/com/nvidia/spark/rapids/tool/planparser/SQLPlanParser.scala
Show resolved
Hide resolved
|
||
// TODO - do we want to use this and rely on stages, but some SQL don't have stages | ||
// so this is less than SQL DF real | ||
val sqlDFWallClockDuration = |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
there is a bug here with the df duration because its looking at supported vs all. I will fix in a followup #5570
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I checked the UI works fine with this PR.
I approve this PR as long as we can followup with PRs addressing the issues listed in #3792
build |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM. The calculation of speedup can be improved in follow on PR's.
build |
not sure why a test was hanging due to this, but fixed. |
build |
fixes #5512 and fixes #5364
Lots of changes here.
So we can try to estimate based on the SQL task times, which is really estimated with the stages task times added up, and we use job overhead To do this, we line up the SQLID to stages and then which operators were in each Stage based on the accumulators. Note not all operators have a mapping to stages.
We take the task time calculations and get a ratio and then try to apply them to the wall clock times to display to the user.
change some of the UDF based operators to have score of 1.2 as we only accelerate the data transfer and udfs still run on cpu.
Each exec checks for udfs and dataset individually to say those aren't supported vs doing the entire SQL operation.
We have 4 levels of recommendation, strongly, recommended, not and not applicable. not applicable is when we have stage or job failures. Need to followup and update UI with good way to display this.
Updated the output to be what we want. Focus on wall clock for the user readable side, the csv file has the task level details. I think we will want to put more in the csv file but for now this works. Outputs the exec level in csv file and outputs the stages used in SQL operations in a csv file.
Need to do more testing on this but would like to get this in and be able to get others to test. A few tests were commented out to update later.
There is also more code cleanup and refactoring I would like to do to make this more readable but that can wait.