논문 인용하기
각 논문마다 생성되어 있는 BibTeX를 사용하시면 자신이 원하는 스타일의 인용 문구를 생성할 수 있습니다.
생성된 BibTeX 코드를 복사하여 BibTeX Parser를 사용해 일반 문자열로 바꾸십시오. 아래의 사이트와 같이 웹에서 변환할 수도 있습니다.
2026
3.
Choi, Jiheon; Oh, Sangyoon
S-CQR: Stratified Calibration for Runtime Prediction in HPC Backfill Scheduling Conference
The 32nd International European Conference on Parallel and Distributed Computing (Euro-Par 2026), 2026.
Links | BibTeX | 태그: Job Scheduling
@conference{jiheon2026scqr,
title = {S-CQR: Stratified Calibration for Runtime Prediction in HPC Backfill Scheduling},
author = {Jiheon Choi and Sangyoon Oh},
url = {https://doi.org/10.1007/978-3-032-35251-4_22},
year = {2026},
date = {2026-08-24},
urldate = {2026-08-24},
booktitle = {The 32nd International European Conference on Parallel and Distributed Computing (Euro-Par 2026)},
keywords = {Job Scheduling},
pubstate = {published},
tppubtype = {conference}
}
2.
Choi, Jiheon; Oh, Sangyoon
The 26th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid 2026), 2026.
Abstract | Links | BibTeX | 태그: Job Scheduling
@conference{CCGrid-2026,
title = {Reducing Backfill Failures with Lightweight Uncertainty Buffers under Workload Drift in HPC Job Scheduling},
author = {Jiheon Choi and Sangyoon Oh},
url = {https://ieeexplore.ieee.org/abstract/document/11618996?casa_token=HPfs4aGNCG8AAAAA:qlQ-AzoiAA2zbe7bMhHlqTmGpTct6zJocgeiYvFowW8aAp8jiP_8-oeswbkEFwBA6FigrNFq42fd},
year = {2026},
date = {2026-05-21},
urldate = {2026-05-21},
booktitle = {The 26th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid 2026)},
abstract = {Backfill scheduling maximizes resource utilization in HPC systems. Typically, conventional schedulers use runtime estimates from user-specified wall-clock limits and prediction models to determine job completion before the next reservation. However, these methods often lead to job termination due to underprediction. As workloads evolve, prediction models degrade, and conventional approaches retrain models when drift is detected. In five HPC traces, we reveal that degradation is concentrated in specific workload segments, indicating unequal underprediction risk across workloads. To tackle this, we present Drift-Adaptive Runtime Prediction (DARP), an adaptation framework that requires minimal retraining. DARP adjusts the buffer to adapt to workload drift, rather than retraining the model. It integrates quantile regression to produce conservative predictions and per-job uncertainty estimates. An adaptive buffer multiplier adjusts based on observed underprediction rates, concentrating protection on high-risk jobs while maintaining efficiency for well-characterized workloads. DARP also provides an immediate response to drift without model retraining. Experiments across five HPC traces show that DARP reduces job termination rates from 58% to under 2%, while eliminating temporal drift and requiring near-zero retraining events compared to conventional methods (average 0.4 vs. 125-1147 for baselines).},
keywords = {Job Scheduling},
pubstate = {published},
tppubtype = {conference}
}
Backfill scheduling maximizes resource utilization in HPC systems. Typically, conventional schedulers use runtime estimates from user-specified wall-clock limits and prediction models to determine job completion before the next reservation. However, these methods often lead to job termination due to underprediction. As workloads evolve, prediction models degrade, and conventional approaches retrain models when drift is detected. In five HPC traces, we reveal that degradation is concentrated in specific workload segments, indicating unequal underprediction risk across workloads. To tackle this, we present Drift-Adaptive Runtime Prediction (DARP), an adaptation framework that requires minimal retraining. DARP adjusts the buffer to adapt to workload drift, rather than retraining the model. It integrates quantile regression to produce conservative predictions and per-job uncertainty estimates. An adaptive buffer multiplier adjusts based on observed underprediction rates, concentrating protection on high-risk jobs while maintaining efficiency for well-characterized workloads. DARP also provides an immediate response to drift without model retraining. Experiments across five HPC traces show that DARP reduces job termination rates from 58% to under 2%, while eliminating temporal drift and requiring near-zero retraining events compared to conventional methods (average 0.4 vs. 125-1147 for baselines).
1.
Choo, Minsol; Oh, Sangyoon
GASched: Goal-Adaptive Hierarchical Reinforcement Learning for Multi-Objective HPC Job Scheduling Conference
The 26th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid 2026), 2026.
Abstract | Links | BibTeX | 태그: Job Scheduling, reinforcement learning
@conference{Choo2026GASched,
title = {GASched: Goal-Adaptive Hierarchical Reinforcement Learning for Multi-Objective HPC Job Scheduling},
author = {Minsol Choo and Sangyoon Oh},
url = {https://ieeexplore.ieee.org/abstract/document/11619039?casa_token=I12UxdEC6WwAAAAA:6uXEo8rxEcsMx4E-92x3dbWIBvbsmM0ihAHoqeFzNmrPqSQTpj4tCiMlm7Hlk_D9AwMnmE5x3P6Z},
year = {2026},
date = {2026-05-21},
urldate = {2026-05-21},
booktitle = {The 26th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid 2026)},
abstract = {High-performance computing (HPC) job scheduling is fundamentally a multi-objective problem because schedulers must simultaneously optimize multiple objectives, such as average slowdown and risk of starvation, facing trade-offs between them. However, which objective matters most can change significantly across workloads and operating conditions, and ignoring these shifts can lead to overall performance degradation. Therefore, understanding and managing the priorities among objectives is a crucial issue for real-world HPC schedulers. However, prior works either target a single objective or treat multi-objective optimization without explicitly capturing their priorities. This oversimplifies the decision problem that real-world schedulers actually face. To address this issue, we present GASched, a Goal-Adaptive HPC Job Scheduler based on hierarchical reinforcement learning (HRL). GASched models whose objectives should be emphasized and allow these priorities to change as the system evolves. A high-level agent selects a subset of objectives to prioritize based on the current queue and cluster status. A lowlevel agent then observes both the current system state and chosen objectives and follows the goal by making scheduling decisions. By repeatedly updating the high-level choices based on long-term feedback, GASched can automatically transfer its focus without manual re-tuning. Through experiments using a diverse set of trace-driven workloads, we demonstrate that GASched balances trade-offs across several key objectives while reducing tail waiting time by up to 93.22% and mean waiting time by up to 89.75% compared with all baseline methods.},
keywords = {Job Scheduling, reinforcement learning},
pubstate = {published},
tppubtype = {conference}
}
High-performance computing (HPC) job scheduling is fundamentally a multi-objective problem because schedulers must simultaneously optimize multiple objectives, such as average slowdown and risk of starvation, facing trade-offs between them. However, which objective matters most can change significantly across workloads and operating conditions, and ignoring these shifts can lead to overall performance degradation. Therefore, understanding and managing the priorities among objectives is a crucial issue for real-world HPC schedulers. However, prior works either target a single objective or treat multi-objective optimization without explicitly capturing their priorities. This oversimplifies the decision problem that real-world schedulers actually face. To address this issue, we present GASched, a Goal-Adaptive HPC Job Scheduler based on hierarchical reinforcement learning (HRL). GASched models whose objectives should be emphasized and allow these priorities to change as the system evolves. A high-level agent selects a subset of objectives to prioritize based on the current queue and cluster status. A lowlevel agent then observes both the current system state and chosen objectives and follows the goal by making scheduling decisions. By repeatedly updating the high-level choices based on long-term feedback, GASched can automatically transfer its focus without manual re-tuning. Through experiments using a diverse set of trace-driven workloads, we demonstrate that GASched balances trade-offs across several key objectives while reducing tail waiting time by up to 93.22% and mean waiting time by up to 89.75% compared with all baseline methods.