A10: Revealing the Power of Neural Networks to Capture
Accurate Job Resource Usage from Unparsed Job Scripts and
Application Inputs
SessionPoster Reception
Author
Event Type
ACM Student Research Competition
Poster
Reception
TimeTuesday, November 14th5:15pm -
7pm
LocationFour Seasons Ballroom
DescriptionNext generation HPC schedulers will rely heavily on
accurate information about resource usage of submitted
jobs. The information provided by users is often
inaccurate and previous prediction models, which rely on
parsed job script features, fail to accurately predict
for all HPC jobs. We propose a new representation of job
scripts and inclusion of application input decks for
resource usage predictions with a neural network. Our
contributions are a method for representing job scripts
as image-like data, an automated method for predicting
job resource usage from job script images and input deck
features, and validation of our methods with real HPC
data. We demonstrate that when job scripts for an
application are very similar, our method performs better
than other methods. We observe an average decrease in
error of 2 node-hours compared to state of the art
methods.
