Quantifying the Carbon Reduction of DAG Workloads: A Job Shop Scheduling Perspective

TL;DR

The paper addresses reducing data-center carbon emissions for batch workloads by recognizing that many jobs decompose into DAG-structured tasks with distinct resource needs. It formalizes a dependency-aware, flexible job-shop scheduling model (FJSP) and evaluates carbon and energy savings under makespan constraints using offline optimization with constraint programming and real carbon-intensity traces. The main contributions show that, on average, carbon emissions can be reduced by up to 25% without extending the optimal makespan, with larger savings achievable when allowing slack (up to ~54%), and that server heterogeneity and job structure significantly influence the achievable reductions. These results provide upper-bound insights to guide the design of online carbon-aware schedulers and reveal the inherent trade-offs between carbon, energy, and makespan in DAG-based batch workloads.

Abstract

Carbon-aware schedulers aim to reduce the operational carbon footprint of data centers by running flexible workloads during periods of low carbon intensity. Most schedulers treat workloads as single monolithic tasks, ignoring that many jobs, like video encoding or offline inference, consist of smaller tasks with specific dependencies and resource needs; however, knowledge of this structure enables opportunities for greater carbon efficiency. We quantify the maximum benefit of a dependency-aware approach for batch workloads. We model the problem as a flexible job-shop scheduling variant and use an offline solver to compute upper bounds on carbon and energy savings. Results show up to $25\%$ lower carbon emissions on average without increasing the optimal makespan (total job completion time) compared to a makespan-only baseline. Although in heterogeneous server setup, these schedules may use more energy than energy-optimal ones. Our results also show that allowing twice the optimal makespan nearly doubles the carbon savings, underscoring the tension between carbon, energy, and makespan. We also highlight key factors such as job structure and server count influence the achievable carbon reductions.

Figures (7)

• Figure 1: A dependency-unaware scheduler attempts to run all tasks in the low-carbon period, but precedence constraints force most tasks to run during a high-carbon period. A dependency-aware scheduler makes a better trade-off for a lower total carbon footprint.
• Figure 2: An example FJSP with homogeneous machines. The top schedule is optimized for the fastest completion time (makespan). The middle schedule is optimized for the lowest carbon footprint, shifting tasks to align with low-carbon periods (bottom) while respecting the makespan constraint.
• Figure 3: For the case of $k=4$ tasks, the dependency structures that we sample from in our experiments.
• Figure 4: Optimal makespan (in epochs, $1$ epoch = 15 min.) for homogeneous and heterogeneous server settings in our experiments. $n = 10$ jobs, $M = 5$ servers, and each job has $k = 4$ operations.
• Figure 5: Carbon savings in the AU-SA grid at different stretch factors $S$. (a) Homogeneous servers. (b) Heterogeneous servers.

Theorems & Definitions (3)

• Definition 2.1: Makespan
• Definition 2.2: Energy Usage
• Definition 2.3: Carbon Emissions