Understand Job Execution in Supercomputers

Elliot Kolker-Hicks (they/them)

Di Zhang

Job Scheduling + Runtime Prediction

• Supercomputers receive lots of jobs (computational tasks) at the same time
• Jobs need to be scheduled to run in optimal order
• This can be very inefficient when we don’t know job runtime
• We can use machine learning to predict job runtime, and improve the scheduling performance

​

​

Metrics for Accuracy

• Paper uses 2 different metrics for measuring prediction accuracy
• Loss - measure of error (y = predicted, x = real)
• Mean Average Error (MAE) -
• E-Loss -

​

​

• Paper is trying to beat the prediction heuristic avgk2 - average runtime last 2 jobs
• avgk2 works very well for predicting runtime, but we can do better
• Our goal is to reimplement the paper, “Improving Backfilling by using Machine Learning to Predict Running Times” from SC’15

Reproducing Prior Work

• Success!
• Achieved similar results to paper
• MAE is actually even better, but E-loss not as good�as good as paper
• We now have a functioning model

Future Plans - Backfilling with Runtime Prediction

• Now that we can reasonably predict runtime, we enter phase 2.
• Does backfilling actually improve with better runtime prediction?
• Use runtime prediction vs requested runtime vs actual runtime in backfilling to compare

Thank You

Q & A