Scientific and enterprise computing grids increasingly host Agentic Large Language Model (LLM) workflows—multi-step, tool-invoking pipelines modelled as Directed Acyclic Graphs (DAGs) of dependent inference tasks. Unlike conventional grid workloads, agentic DAGs exhibit stochastic node weights and unbounded retry sub-graphs, creating severe carbon tail-risk under volatile electricity markets. Existing grid schedulers optimise expected makespan and energy but provide no guarantee over worst-case carbon expenditure. We present MacroCarbon, a three-layer Grid Resource Broker combining: (i) a reverse Mixed Data Sampling Quantile Regression (MIDAS-QR) carbon-intensity forecaster driven by low-frequency macroeconomic covariates (EU natural gas futures, carbon allowance prices, and per-country industrial production indices); (ii) a spatial site-selection layer routing each DAG to the geo-distributed site with the minimum forecasted Carbon Conditional Value-at-Risk (C-CVaR); and (iii) an intra-site elastic provisioner enforcing C-CVaR constraints by gating maximum retry depth <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$K_{\max }$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>K</mi> <mo movablelimits="true">max</mo> </msub> </math></EquationSource> </InlineEquation> and cascading tasks to smaller GPU partitions. Evaluated in CloudSim Plus 3.0 over four European grid sites using real 2021–2026 ElectricityMap carbon-intensity traces and SWE-bench-lite agentic DAG workloads (including a held-out 2025 Q1 generalisation test), MacroCarbon achieves 53.9% carbon-CVaR suppression versus a static-budget baseline and 18.4% forecasting-error reduction over a standard time-series baseline (18.7% on the held-out epoch, confirming out-of-sample generalisation). Our primary, capability-matched evaluation equips single-site baselines (MOHEFT, EcoServe) with a common spatial wrapper, separating the effect of spatial capability from MacroCarbon’s algorithm: under a legacy batch arrival trace these fairly-equipped baselines are competitive, but under realistic agentic bursty/diurnal arrivals MacroCarbon reduces per-job carbon CVaR by 45–51% over them (Wilcoxon <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$p{<}0.01$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo><</mo> <mn>0.01</mn> </mrow> </math></EquationSource> </InlineEquation>) while sustaining 86–88% makespan SLA. The advantage is therefore tail-risk control under realistic agentic load—not spatial routing per se—and it is robust to correlated cross-site carbon intensity.

MacroCarbon: A CVaR-Constrained Grid Resource Broker for Agentic LLM DAG Workflows on Geo-Distributed Heterogeneous Computing Grids

期刊论文
发表日期：

2026年7月30日

查看全文

Hemant P,
Afreen Arif

摘要

Scientific and enterprise computing grids increasingly host Agentic Large Language Model (LLM) workflows—multi-step, tool-invoking pipelines modelled as Directed Acyclic Graphs (DAGs) of dependent inference tasks. Unlike conventional grid workloads, agentic DAGs exhibit stochastic node weights and unbounded retry sub-graphs, creating severe carbon tail-risk under volatile electricity markets. Existing grid schedulers optimise expected makespan and energy but provide no guarantee over worst-case carbon expenditure. We present MacroCarbon, a three-layer Grid Resource Broker combining: (i) a reverse Mixed Data Sampling Quantile Regression (MIDAS-QR) carbon-intensity forecaster driven by low-frequency macroeconomic covariates (EU natural gas futures, carbon allowance prices, and per-country industrial production indices); (ii) a spatial site-selection layer routing each DAG to the geo-distributed site with the minimum forecasted Carbon Conditional Value-at-Risk (C-CVaR); and (iii) an intra-site elastic provisioner enforcing C-CVaR constraints by gating maximum retry depth $K_{\max }$ $K_{max}$ and cascading tasks to smaller GPU partitions. Evaluated in CloudSim Plus 3.0 over four European grid sites using real 2021–2026 ElectricityMap carbon-intensity traces and SWE-bench-lite agentic DAG workloads (including a held-out 2025 Q1 generalisation test), MacroCarbon achieves 53.9% carbon-CVaR suppression versus a static-budget baseline and 18.4% forecasting-error reduction over a standard time-series baseline (18.7% on the held-out epoch, confirming out-of-sample generalisation). Our primary, capability-matched evaluation equips single-site baselines (MOHEFT, EcoServe) with a common spatial wrapper, separating the effect of spatial capability from MacroCarbon’s algorithm: under a legacy batch arrival trace these fairly-equipped baselines are competitive, but under realistic agentic bursty/diurnal arrivals MacroCarbon reduces per-job carbon CVaR by 45–51% over them (Wilcoxon $p{<}0.01$ $p < 0.01$ ) while sustaining 86–88% makespan SLA. The advantage is therefore tail-risk control under realistic agentic load—not spatial routing per se—and it is robust to correlated cross-site carbon intensity.