智能驾驶场景智能算力卸载数据集

在智能驾驶场景中，车辆或者路边感知设备会实时收集数据，这些数据需要在极短的时间内得到高效处理，进而支撑后续的智能驾驶决策。这些任务通常被建模为有向无环图（DAG），其子任务间存在复杂的依赖关系，对调度算法提出了极高要求。尽管已有如Google Cluster Data 2011这样的真实轨迹数据，为研究提供了宝贵的输入，但缺乏与之配套的、标准化的算法输出性能基准，使得不同调度策略之间的公平对比和深度评估变得困难。其次，这些数据没有体现出强化学习等任务调度算法的调度和处理效果，缺少对任务调度适配性和性能的描述，使得准确分析任务的调度优化空间和调度算法的调度性能较为困难。本数据集旨在填补这一空白。我们以集群数据为真实工作负载基础，通过一个高度可配置的离散事件调度仿真器，生成了多轮次、大规模的DAG调度实验数据。它不仅包含了仿真的最终性能结果（以DAG完成时间为关键指标），还记录了详细的作业执行轨迹，从而形成了一个从“问题输入”（真实任务事件）到“解决方案输出”（调度器性能）的完整数据链。这套数据将为调度算法研究者提供一个端到端的、可复现的基准测试环境，极大地促进调度优化领域的创新与发展。数据集共包括2个数据文件，其中：(1) task_trace.csv 是任务执行数据记录，以行为单位记录了每个任务每一次状态变化的事件。包含时间戳、任务ID、机器ID、事件类型等13个字段，数据量是260MB；(2) dag_completion_times.txt是多轮次任务完成时间日志，数据量1.3MB。In autonomous driving scenarios, vehicles or roadside perception devices collect data in real time. Such data requires efficient processing within an extremely short time frame to support subsequent autonomous driving decision-making. These tasks are usually modeled as Directed Acyclic Graphs (DAGs), where complex dependencies exist among subtasks, posing extremely high requirements for scheduling algorithms. Although real-world trajectory datasets such as Google Cluster Data 2011 have provided valuable inputs for research, there is a lack of supporting standardized performance benchmarks for algorithm outputs, which makes fair comparison and in-depth evaluation of different scheduling strategies difficult. Furthermore, these datasets do not reflect the scheduling and processing effects of task scheduling algorithms such as Reinforcement Learning (RL), nor do they describe the adaptability and performance of task scheduling, making it difficult to accurately analyze the scheduling optimization space of tasks and the scheduling performance of algorithms. This dataset aims to fill this gap. We take cluster data as the basis for real-world workloads, and generate multi-round, large-scale DAG scheduling experimental data through a highly configurable discrete-event scheduling simulator. It not only includes the final simulation performance results (with DAG completion time as the core metric), but also records detailed job execution traces, forming a complete data pipeline from "problem input" (real task events) to "solution output" (scheduler performance). This dataset will provide scheduling algorithm researchers with an end-to-end, reproducible benchmarking environment, greatly promoting innovation and development in the field of scheduling optimization. The dataset consists of 2 data files in total: (1) task_trace.csv: this is a task execution data record, which logs each state change event of each task on a per-line basis. It contains 13 fields such as timestamp, task ID, machine ID, and event type, with a total size of 260 MB; (2) dag_completion_times.txt: this is a multi-round task completion time log, with a total size of 1.3 MB.

搜集汇总

背景与挑战

背景概述

该数据集针对智能驾驶场景中任务调度缺乏标准化性能基准的挑战，以真实集群数据为基础，通过可配置仿真器生成了包含任务执行轨迹和完成时间的大规模实验数据。它为调度算法研究提供了端到端、可复现的基准测试环境，旨在促进该领域的创新与发展。