Differentiable Combinatorial Scheduling at Scale

Authors: Mingju Liu, Yingjie Li, Jiaqi Yin, Zhiru Zhang, Cunxi Yu

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comparative evaluations on both synthetic and real-world benchmarks highlight our capability to significantly improve the optimization efficiency of scheduling, surpassing state-of-the-art solutions offered by commercial and open-source solvers such as CPLEX, Gurobi, and CP-SAT in the majority of the designs. Our experimental results demonstrate significant improvements in optimization efficiency over state-of-the-art (SOTA) methods solved with commercial solvers CPLEX (IBM, 2023), Gurobi (Gurobi Optimization, LLC, 2023) and open-source CP-SAT solver (Perron & Didier; Perron et al., 2023).
Researcher Affiliation Academia 1University of Maryland, College Park 2Cornell University.
Pseudocode Yes Algorithm 1 Differentiable Scheduling
Open Source Code Yes Our experimental setups and implementations are available at https://github.com/Yu-Maryland/ Differentiable_Scheduler_ICML24.
Open Datasets Yes Our GPU workloads/graphs are derived using six designs from the EPFL Benchmark Suite (Amar u et al., 2015), alongside baseline SDC+LP formulation solved by the SOTA commercial solvers, CPLEX (IBM, 2023) and Gurobi (Gurobi Optimization, LLC, 2023), as well as open-source tool CP-SAT solver (Perron & Didier).
Dataset Splits No The paper does not explicitly provide training, validation, or test dataset splits. It describes optimization against synthetic and benchmark designs rather than typical data splits for supervised learning.
Hardware Specification Yes All experiments were conducted using an an Intel Xeon Gold 6418H CPU and NVIDIA RTX 4090 GPU.
Software Dependencies Yes Our experimental results demonstrate significant improvements in optimization efficiency over state-of-the-art (SOTA) methods solved with commercial solvers CPLEX (IBM, 2023), Gurobi (Gurobi Optimization, LLC, 2023) and open-source CP-SAT solver (Perron & Didier; Perron et al., 2023).
Experiment Setup Yes We predefined the ratio of LP and the factor of our method, setting R = λ = 100 for all EPFL designs and R = λ = 10 for all synthetic workloads, respectively. We set the targeted latency to be L = 10 for the experiments. An initial instability observed during the first 10 epochs can be attributed to the necessity of a warm-up phase for the Gumbel-Softmax mechanism employed in discrete sampling by our method. Furthermore, we explored the impact of different weight decay settings by transitioning from the Adam optimizer to Adam W.