1300-GRP From Forecasts to Decisions: Adaptive Container Resource Allocation using Forecast-Guided Reinforcement Learning

Presenter Information

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Container schedulers reserve CPU and memory before a job begins, relying on either user-specified resource requests or demand forecasts optimized for prediction accuracy. However, inaccurate estimates can be costly. Too little CPU throttles the job, too little memory kills it, and too much leaves capacity idle while other jobs wait. We keep an LSTM forecast of each job's peak demand and learn the safety margin on top of it from the system cost it causes, not from forecast error. A grid search picks the starting margin and Reinforcement Learning (RL) fine-tunes it in a cluster simulator driven by the Alibaba 2018 trace. Our results show that LSTM-RL cuts OOM kills from 21.9% to 1.2%, SLA violations from 31% to 13%, and is 15–44% cheaper than user requests.

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

1300-GRP From Forecasts to Decisions: Adaptive Container Resource Allocation using Forecast-Guided Reinforcement Learning

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Container schedulers reserve CPU and memory before a job begins, relying on either user-specified resource requests or demand forecasts optimized for prediction accuracy. However, inaccurate estimates can be costly. Too little CPU throttles the job, too little memory kills it, and too much leaves capacity idle while other jobs wait. We keep an LSTM forecast of each job's peak demand and learn the safety margin on top of it from the system cost it causes, not from forecast error. A grid search picks the starting margin and Reinforcement Learning (RL) fine-tunes it in a cluster simulator driven by the Alibaba 2018 trace. Our results show that LSTM-RL cuts OOM kills from 21.9% to 1.2%, SLA violations from 31% to 13%, and is 15–44% cheaper than user requests.