Green or Fast? Learning to Balance Cold Starts and Idle Carbon in Serverless Computing

Bowen Sun; Christos D. Antonopoulos; Evgenia Smirni; Bin Ren; Nikolaos Bellas; Spyros Lalis

cs.DCFeb 27, 2026

Green or Fast? Learning to Balance Cold Starts and Idle Carbon in Serverless Computing

Authors: Bowen Sun, Christos D. Antonopoulos, Evgenia Smirni, Bin Ren, Nikolaos Bellas, Spyros Lalis

Organizations: Department of Computer Science, William & Mary, Williamsburg, VA, USA · Department of Computer Science, University of Thessaly, Volos, Greece

Abstract

Serverless computing simplifies cloud deployment but introduces new challenges in managing service latency and carbon emissions. Reducing cold-start latency requires retaining warm function instances, while minimizing carbon emissions favors reclaiming idle resources. This balance is further complicated by time-varying grid carbon intensity and varying workload patterns, under which static keep-alive policies are inefficient. We present LACE-RL, a latency-aware and carbon-efficient management framework that formulates serverless pod retention as a sequential decision problem. LACE-RL uses deep reinforcement learning to dynamically tune keep-alive durations, jointly modeling cold-start probability, function-specific latency costs, and real-time carbon intensity. Using the Huawei Public Cloud Trace, we show that LACE-RL reduces cold starts by 51.69% and idle keep-alive carbon emissions by 77.08% compared to Huawei's static policy, while achieving better latency-carbon trade-offs than state-of-the-art heuristic and single-objective baselines, approaching Oracle performance.

Figures & tables

Dataset Source	Description	Usage in LACE-RL Framework
Cold Start Logs	31-day cold start traces with breakdowns of pod allocation cost , code deploy cost .	Used to construct function-specific cold start latency ( $L_{cs}$ ) by runtime
Request-Level Logs	Per-invocation metadata including timestamp , execution time , podID , CPU , and memory request	Used to simulate invocation traces, calculate pod reuse intervals, and model pod reuse probabilities
Runtime and Trigger Metadata	Static table mapping each function with its runtime language and trigger type	Used to determine runtime-sensitive latency and filter function categories

TABLE I: Summary of Huawei Dataset Components Used in the LACE-RL Framework

Panel of Figure 2 — Fig. 1: Characterization of the Huawei Cloud Trace [ 5 ] .

Fig. 2: Impact of keep-alive timeout for two different but representative functions in the Huawei dataset. In both cases, longer timeouts reduce cold starts but increase idle carbon footprint. Depending on the function, idle carbon may even surpass (significantly) the carbon for execution (right plot).

Panel of Figure 4 — Fig. 3: Observed variability in carbon intensity of the electrical grid [ 14 ] (a) and Huawei’s function memory footprint (b).

Fig. 4: LACE-RL overview.

Benchmark Function Name	Input Size / Dimension	Memory Consumption	Cold Start Time (ms)	Compute Time (ms)	Cold-start Active Energy (J)	Compute-phase Active Energy (J)	1-min Keep-alive-phase Active Energy (J)	Per-pod Compute-Phase Total Power (W)	Per-pod Keep-alive-phase Total Power (W)	Lambda Idle Total Power Ratio (Keep-alive/Compute)
Float Operations	10,000,000	44 MB	112.2	3340.86	0.94	15.08	78.29	6.37	3.19	0.50
MatMul	10,000	95 MB	166.5	2393.41	0.27	144.41	76.98	86.64	28.89	0.33
Linpack	100,000	97 MB	76.33	6401.45	0.7	436.9	92.4	147.29	70.82	0.48
Image Processing	28.4 MB	68 MB	2441.68	6761.82	11.13	20.69	81.6	4.98	3.21	0.64
Video Processing	742 KB	233 MB	12414.77	2403.04	19.05	6.82	72.68	4.65	3.03	0.65
Chameleon	[500,100]	57 MB	71.6	249.52	0.52	1.84	81.1	9.27	3.14	0.34

TABLE II: Energy profiling of serverless function pods under cold start, compute, and keep-alive phases.

Panel of Figure 7 — Fig. 5: General testing workload results across key metrics. The dashed line shows the optimal value.

Fig. 6: For the General workload LACE-RL achieves the best tradeoff between cold starts and keep-alive carbon.

Panel of Figure 9 — Fig. 7: Comparison based on the composite metrics (lower is better). LACE-RL always outperforms the other strategies.

Panel of Figure 10 — Fig. 8: Long-tailed workload results: the dashed line shows the optimal value.

Fig. 9: Long-tailed workload: LACE-RL achieves the most balanced tradeoff between cold starts and carbon emissions.

Case	Metrics	Oracle	LACE-RL	Degradation
General	Keep-alive Carbon (gCO 2 )	2.88	3.058	6.181%
General	Cold Start Count (Times)	1069	1146	7.203%
Long-tailed	Keep-alive Carbon (gCO 2 )	3.9	4.25	8.974%
Long-tailed	Cold Start Count (Times)	807	897	11.152%

TABLE III: Comparison of LACE-RL vs. Oracle

Panel of Figure 13 — Fig. 10: Sensitivity of LACE-RL to tunable parameters (a) and interpretability (b).

Cloud Computing

Cloud computing studies on-demand delivery and management of compute, storage, and services across distributed infrastructure.

See all 47 papers

Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control
Apr 7, 2026Zeyu Wang, Cuiqianhe Du, Renyue Zhang +3Cloud Computing
University of California, Los Angeles, Los Angeles, USA · University of California, Berkeley, Berkeley, USA · New York University, New York, USA +3
SLA-Constrained Carbon-Aware Routing in Geo-Distributed Serverless Clouds
Jul 24, 2026Anmol Chaudhary, Rahul MishraCloud Computing
Department of Electronics and Computer Engineering, NIAMT Ranchi · National Institute of Advanced Manufacturing Technology Ranchi, India
GAR: Carbon-Aware Routing for LLM Inference via Constrained Optimization
May 12, 2026Disha Sheshanarayana, Rajat Subhra Pal, Manjira Sinha +1LLM Inference Efficiency Constrained Optimization
Manipal University Jaipur · TCS Research

Apr 7, 2026J/K move · Enter open · S save

Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control

Zeyu Wang, Cuiqianhe Du, Renyue Zhang, Kejian Tong+2

University of California, Los Angeles, Los Angeles, USA · University of California, Berkeley, Berkeley, USA · New York University, New York, USA+3

Cloud Computing

Open paper AskPDF arXiv

Serverless computing eliminates infrastructure management overhead but introduces significant challenges regarding cold start latency and resource utilization. Traditional static resource allocation often leads to inefficiencies under variable workloads, resulting in performance degradation or excessive costs. This paper presents an adaptive engineering framework that optimizes serverless performance through event-driven architecture and probabilistic modeling. We propose a dual-strategy mechanism that dynamically adjusts idle durations and employs an intelligent request waiting strategy based on slot survival predictions. By leveraging sliding window aggregation and asynchronous processing, our system proactively manages resource lifecycles. Experimental results show that our approach reduces cold starts by up to 51.2% and improves cost-efficiency by nearly 2x compared to baseline methods in multi-cloud environments.

Abstract

Cloud Computing

Explore similar work

Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control

SLA-Constrained Carbon-Aware Routing in Geo-Distributed Serverless Clouds

GAR: Carbon-Aware Routing for LLM Inference via Constrained Optimization