AI-Guided Decision Making: Use case: Scheduling

Arunavo Dey, Tanzima Islam

Texas State University

Application Fingerprinting Update

• Forecast power usage based on job description and rank jobs
• Extract common power usage patterns across jobs
• [TODO] Build a signature for a cluster of jobs with similar compositional power usage patterns
• [TODO] Integrate these signatures in the AI model for better decision-making about scheduling

​

Build AI/ML Methodology to Guide the Scheduler

Attention-based Autoencoder

Sequence-based Models

Attention-based Autoencoder

New job description

Rank Job

Scheduler

Offline training

Online decision-making

New samples with power measurements

Adaptation

Test-time adaptation with streaming measurements can update the model without rebuilding from scratch

Changes in the Job Dictionary

Changes in assign_to_job

Changes in the schedule Function

Changes in Policies

Changes in Configurations

Job Throughput

Runtime in seconds

Backup Slides

Changes in RAPS

Old function in dataloader

Changes in RAPS

Changed function in dataloader

Changes in RAPS

Changed function in dataloader

Changes in RAPS

Old function in dataloader ( load data from df )

Changes in RAPS

Changed function in dataloader ( load data from df )