Open Access

ARTICLE

CASH: Confidence-Calibrated Deep Feature Crossing for Classification-Aware Heterogeneous Task Scheduling

Chuanlin Jian¹, Yuanchen Sun², Xiangcheng Liu¹, Xuming Huang³, Samaneh Beheshti Kashi⁴, Xing Hu^1,*

1 School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, China
2 School of Artificial Intelligence, Shenyang Normal University, Shenyang, China
3 Department of Computer Sciences, University of Wisconsin–Madison, 1210 W Dayton Street, Madison, WI, USA
4 Zhejiang Engineering Research Center of Interventional Medicine Engineering and Biotechnology, The Fifth Affiliated Hospital of Wenzhou Medical University, Lishui, China

* Corresponding Author: Xing Hu. Email:

Computers, Materials & Continua 2026, 89(2), 29 https://doi.org/10.32604/cmc.2026.086441

Received 30 May 2026; Accepted 22 July 2026; Issue published 15 September 2026

Abstract

Heterogeneous clusters now carry most artificial-intelligence, scientific-computing, cloud, and edge-assisted workloads, yet scheduling them well remains difficult: workload semantics, hardware capability, queue state, and energy behavior are tightly coupled. Existing schedulers, whether heuristic, learning-based, or built on reinforcement learning, tend to rely on coarse resource requests, overlook the compatibility between task types and node classes, or carry a heavy training and deployment cost. This paper presents Classification-Aware Scheduling for Heterogeneous clusters (CASH), a confidence-calibrated, classification-aware scheduling framework that integrates workload profiling, deep feature crossing, gradient-boosted boundary modeling, and risk-aware heterogeneous resource mapping. CASH constructs a multi-view workload representation from resource requests, submission context, queue attributes, and historical user behavior. A Deep & Cross Network captures bounded-degree high-order feature interactions; an eXtreme Gradient Boosting (XGBoost) branch models the sharp decision boundaries typical of tabular scheduling logs. A validation-calibrated, confidence-aware fusion rule combines the two branches into task-type probabilities with attached uncertainty estimates. The scheduler then performs class-constrained candidate selection and chooses execution nodes by minimizing a composite score over predicted execution time, energy, load imbalance, semantic mismatch, and prediction uncertainty. In trace-driven simulations on Massachusetts Institute of Technology (MIT) Supercloud logs, CASH improves resource matching, response time, makespan, and energy consumption over representative baselines while keeping load balance competitive. The results indicate that an explicit, calibrated link between workload semantics and hardware capability is a practical basis for efficient and interpretable cluster scheduling.

---

Keywords

---