Neural DAG Scheduling via One-Shot Priority Sampling
Authors: Wonseok Jeon, Mukul Gagrani, Burak Bartan, Weiliang Will Zeng, Harris Teague, Piero Zappi, Christopher Lott
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically show that our algorithm generates better schedules than both non-neural and neural baselines across various real-world and synthetic scheduling tasks. |
| Researcher Affiliation | Industry | Qualcomm AI Research |
| Pseudocode | Yes | Algorithm 1 Neural DAG Scheduler via One-Shot Priority Sampling |
| Open Source Code | No | The paper does not provide a direct statement or link to their own open-source code for the described methodology. It mentions code used for baselines or third-party libraries but not their own implementation. |
| Open Datasets | Yes | We evaluate our model on randomly generated JSSP instances from Zhang et al. (2020) which are defined by number of jobs Nj and number of machines Nm. We use Wang et al. s TPC-50/TPC-100/TPC-150 datasets for our experiments. |
| Dataset Splits | Yes | We train our algorithm for 100 training graphs and evaluate it for 50 unseen test graphs. We use the first 100 instances as training instances and the last 50 instances as test instances. The dataset for each of TPC-50, TPC-100, TPC-150 experiments consists of 50 training graphs and 10 test graphs. For each synthetic graph distribution, we consider the graph size equal to either 500 or 1000 and generate a training set of 3000 graphs and 300 unseen test graphs. |
| Hardware Specification | Yes | For all the tasks, we use a machine with a single GPU (Nvidia Tesla V-100) with 32 GB memory that is also used to train and evaluate our model. |
| Software Dependencies | No | The paper mentions software like 'Adam optimizer', 'Google OR-Tools', and 'CP-SAT solver' but does not provide specific version numbers for these or other software dependencies necessary for reproducibility. |
| Experiment Setup | Yes | We use clogits = 0.001 and ϵ = 0.1 for clipping the denominator in Eq. (13). We train our model for 20 epochs with 5 random seeds and pick the best performing model. We use the number of samples N = 1000. |