Aceso: Latency-Aware Geo-Distribution of Microservices for Cost-Effective Carbon Reduction
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Aceso: Latency-Aware Geo-Distribution of Microservices for Cost-Effective Carbon Reduction".
Jane: Here's the gist of the paper. Microservices are dominant in cloud computing but pose complex deployment challenges that require carbon-aware scheduling to mitigate their significant energy consumption HuntMS presents an adaptive placement system that jointly optimizes carbon and cost while respecting latency constraints.
Tom: Problem Motivation The paper addresses the challenge of deploying microservices in geographically constrained infrastructures for SMEs, which often lack access to global-scale infrastructure.
Jane: Existing carbon-aware solutions assume global infrastructure, which is not true for many regional SMEs with microservice applications.
Tom: The core question explored is how SMEs can deploy microservice applications in a carbon-efficient and cost-effective manner within geographically constrained infrastructures while maintaining application performance.
Title and authors: Tom: So the HuntMS system is an adaptive carbon-and-efficiency-aware microservice placement system that jointly optimizes carbon emissions and monetary cost while respecting latency constraints. It uses a scalable genetic algorithm for placement identification across regions with different carbon intensities.
Jane: They are saying this system helps SMEs deploy microservice applications in a carbon-efficient and cost-effective manner within geographically constrained infrastructures while maintaining application performance. That’s the main question they're trying to answer, focusing on that tricky balance.
Lu: The system works by evolving a population of candidate placements over time using that genetic algorithm. It's designed to rapidly find near-optimal placements across regions that have heterogeneous carbon intensity, which means it can handle those different electricity mixes.
Meng: What about the real-time aspect? If a service suddenly gets much more traffic than expected, or if the local carbon intensity spikes unexpectedly because of grid changes, how quickly does this system react? Is it slow to adapt?
Lalam: It’s designed to be event-driven. The system can re-invoke the Placement Optimizer when triggered by either carbon or workload changes, reacting to forecasted traffic shifts and carbon intensity deviations. This gives it a way to stay stable under dynamic conditions without needing constant recalculation of the entire thing.
Tom: Event-driven optimization sounds much better for real-world deployment than just running a big optimization job once and forgetting about it. It means it can react quickly when things change on the ground, which is crucial for dynamic environments.
Jane: That adaptability is key, especially since carbon intensity isn't static; it changes hourly based on local energy mix, which they emphasize in their data visualization. So the system needs to be fluid to keep up with that variance in real time.
Lu: They also mention how they reduce the search space using two specific techniques: microservice pinning and region filtering based on a Pareto-style rule. The pruning rule keeps regions only if they do not worsen all metrics compared to the base region.
Meng: That Pareto-style rule is where I get cautious. If a region improves one metric but makes both carbon and cost worse compared to the base region, is that really worth keeping it in the search? It sounds like they are being very strict about keeping things that offer a clear advantage across the board.
The paper's summary: Tom: They suggest improvements by focusing on how they reduce the search space dramatically. One key improvement is microservice pinning, which pins inherently immovable services like frontends and databases, making them stay put.
Jane: Pinning those things down seems very practical because you don't want to move your core data store just because it’s a little greener somewhere else. That simplifies the problem significantly for the rest of the system by removing unnecessary variables from consideration.
Lu: They also suggest ranking microservices based on their activation order to pin later activated ones when end-to-end latency approaches the service level objective. That's smart; you only try to move things that are close to hitting your performance limits anyway.
Meng: If you pin services based on activation order, how do you know what the activation order is going to be? That sounds like another piece of data they have to trust for their pruning strategy. What if the traffic pattern changes completely?
Lalam: They also suggest focusing on offloading microservice subtrees that are activated early in a request execution and do not belong to the critical path. This eliminates decisions that are infeasible or low-impact, which reduces the optimization overhead while maintaining latency guarantees.
Tom: So they’re trying to be surgical about where they apply these optimizations, only bothering with services that matter for performance or cost reduction. That sounds like a good way to keep the system from getting bogged down in unnecessary complexity.
Jane: It sounds like the improvement is less about finding a perfect global solution and more about making the search process incredibly efficient so it can find a very good solution quickly enough for SMEs. That’s what makes it feasible for real business use.
The paper's improvements: Tom: So to wrap up, Aceso enables carbon- and cost-aware microservice deployment for latency-sensitive applications in regionally constrained infrastructures. It shows that you can use this adaptive placement system to achieve significant reductions in both emissions and operational costs compared to a static deployment within a single country.
Jane: The results they showed are pretty compelling: HuntMS quickly adapts to real-time changes in workload and carbon intensity, reducing carbon emissions by thirty-seven point four percent and operational cost by three point six percent, on average, compared to a static deployment within one country. Plus, they kept the average end-to-end latency around ninety-three point eight milliseconds with only violations of two percent in the European setting <ref:2603.10768#pg1>.
Lu: The comparison against baselines like Caribou and Vanilla GA shows that HuntMS achieves the best practical carbon savings while maintaining latency close to the optimal linear programming solution. That’s a solid performance indicator for an optimization approach.
Meng: From a deployment standpoint, if you look at the scalability, HuntMS converges in under two point two minutes even with one thousand microservices and all global regions <ref:2603.10768#pg1>. That convergence time is much faster than what I'd expect from a complex optimization problem when you scale up that way.
Lalam: The system’s event-driven optimization allows it to react to workload changes, demonstrating stability compared to reactive modes without forecasting. It basically stays stable under dynamic conditions by reacting immediately when something shifts.
Tom: So the main implication for SMEs is that they don't have to choose between being environmentally responsible and being cost-effective; this paper shows they can do both with HuntMS because it handles the latency constraints dynamically.
Jane: It seems like a practical path towards sustainable microservice deployment for the real-world majority of SMEs, which is where this research is truly useful. It’s about making complex infrastructure decisions manageable for businesses operating in real, limited environments.
Conclusion: Tom: So we've looked at Aceso, which is about latency-aware geo-distribution of microservices for cost-effective carbon reduction. Basically, this system uses a genetic algorithm to place services where they are both cheap and green while keeping performance tight.
Jane: That’s the core idea: making sure Small and Medium Enterprises can deploy these microservices sustainably without breaking their service level objectives. It's about balancing real-world operational limits with environmental responsibility.
Lu: I think what really stands out is how they handle that huge search space by using insight-driven pruning techniques, specifically microservice pinning and Pareto-style region filtering to cut down the variables they have to consider early on.
Meng: But I still wonder about the real deployment cost of running that optimization every time a new service starts up. If you're shipping models, how much overhead does this adaptive placement add to your infrastructure management?
Lalam: They tackle that by making it event-driven; the system re-invokes the optimizer when carbon or workload changes happen, which keeps things stable without needing constant recalculation of everything. I'm quite sure that dynamic reaction is what makes it practical for running production models.
Tom: Event-driven sounds way better than just running one big optimization job and forgetting about it. It means the system can react quickly when things change on the ground.
Jane: That adaptability is key because carbon intensity isn't static; it changes hourly based on local energy mix, so the system has to be fluid to keep up with that variance.
Lu: The paper also points out that they focus offloading microservice subtrees activated early in a request execution and not belonging to the critical path, which eliminates decisions that are infeasible or low-impact. That reduces optimization overhead while maintaining latency guarantees.
Meng: So they’re trying to be surgical about where they apply these optimizations, only bothering with services that matter for performance or cost reduction. That sounds like a good way to keep the system from getting bogged down in unnecessary complexity when deploying models at scale.
Tom: To wrap up, Aceso shows you can use this adaptive placement system to achieve significant reductions in both emissions and operational costs compared to a static deployment within a single country.
Jane: The results they showed are pretty compelling: HuntMS quickly adapts to real-time changes in workload and carbon intensity, reducing carbon emissions by thirty-seven point four percent and operational cost by three point six percent, on average, compared to a static deployment within one country. Plus, they kept the average end-to-end latency around ninety-three point eight milliseconds with only violations of two percent in the European setting <ref:2603.10768#pg1>.
Lu: The comparison against baselines like Caribou and Vanilla GA shows that Aceso achieves the best practical carbon savings while maintaining latency close to the optimal linear programming solution. That’s a solid performance indicator for an optimization approach.
Meng: From a deployment standpoint, if you look at the scalability, Aceso converges in under two point two minutes even with one thousand microservices and all global regions <ref:2603.10768#pg1>. That convergence time is much faster than what I'd expect from a complex optimization problem when you scale up that way.
Lalam: The system’s event-driven optimization allows it to react to workload changes, demonstrating stability compared to reactive modes without forecasting. It basically stays stable under dynamic conditions by reacting immediately when something shifts.
Tom: So the main implication for SMEs is that they don't have to choose between being environmentally responsible and being cost-effective; this paper shows they can do both with Aceso because it handles the latency constraints dynamically.
Jane: It seems like a practical path towards sustainable microservice deployment for the real-world majority of SMEs, which is where this research is truly useful. It’s about making complex infrastructure decisions manageable for businesses operating in real, limited environments.
Tom: We've covered Aceso today, but if you want to see how these placement ideas apply to something completely different—like those distance oracles we talked about—we'll be looking at that next paper.
Georgia Christofidi, Francisco Álvarez-Terribas, Ioannis Roumpos, Nicolas Kourtellis, Jesus Omaña Iglesias, Thaleia Dimitra Doudali
IMDEA Software Institute · Telefonica Research
cs.DC
Submitted: 2026-03-11
Updated: 2026-10-29
Code: https://github.com/delimitrou/DeathStarBench
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Microservices are dominant in cloud computing but pose complex deployment challenges that require carbon-aware scheduling to mitigate their significant energy consumption HuntMS presents an adaptive…
Key concepts
- Multi-Objective Optimization
- This is the core problem HuntMS solves by trying to improve three competing goals simultaneously: reducing carbon footprint, minimizing monetary cost, and ensuring low latency. The system must find a placement that satisfies all these conditions as best as possible.
- Genetic Algorithm
- HuntMS uses a scalable genetic algorithm within its Placement Optimizer. This technique evolves a set of potential service placements over time by mimicking natural selection to find the best overall configuration that balances carbon, cost, and latency constraints.
- Insight-Driven Pruning
- To speed up the search process, HuntMS intelligently reduces the number of regions it considers. It uses rules like microservice pinning (fixing essential services) and Pareto-style filtering to discard regions that offer no significant improvement across all three objectives compared to a baseline.
- Event-Driven Optimization
- The system reacts dynamically rather than running constantly. It re-evaluates the placement strategy only when triggered by specific events, such as changes in workload or carbon intensity, allowing it to adapt quickly and efficiently without unnecessary computation.
Terminology
Summary
Microservices are dominant in cloud computing but pose complex deployment challenges that require carbon-aware scheduling to mitigate their significant energy consumption HuntMS presents an adaptive placement system that jointly optimizes carbon and cost while respecting latency constraints. p. 1 Therefore, this
paper addresses the following question:
How can SMEs deploy microservice applications in a carbon-efficient
and cost-effective manner within geographically constrained infrastructures, while maintaining application performance? To answer this question, we present HuntMS: an adaptive carbonand efficiency-aware microservice placement system that jointly
optimizes carbon emissions and monetary cost, while respecting
latency constraints.Where the paper says it
Problem Motivation
The paper addresses the challenge of deploying microservices in geographically constrained infrastructures for SMEs, which often lack access to global-scale infrastructure. Existing carbon-aware solutions assume global infrastructure, which is not true for many regional SMEs with microservice applications. p. 0 However, most existing solutions target batch,
high-performance, or serverless workloads and assume access to
global-scale infrastructure. Such an assumption does not hold for
many national or regional small to medium-sized enterprises (SMEs)
with microservice applications, which represent the real-world majority. For SMEs with limited infrastructure flexibility, this is both a challenge (operating
under strict latency, regulatory, and resource constraints) and an
opportunity to adopt carbon-aware, cost-efficient, regionally deployable solutions that support global climate goals. Therefore, this
paper addresses the following question:
How can SMEs deploy microservice applications in a carbon-efficient
and cost-effective manner within geographically constrained infrastructures, while maintaining application performance?Where the paper says it
Where the paper says it
System Design
HuntMS is designed to solve a constrained multi-objective optimization problem balancing three key objectives: Carbon Efficiency, Cost Efficiency, and Latency Compliance. The system consists of four components: the Traffic Forecaster, the Workload Profiler, the Placement Optimizer, and the Scheduler. p. 4 These organizations often
face limited resource availability, strict data sovereignty regulations,
and heterogeneous resource capacities, making traditional carbonaware strategies impractical. HuntMS addresses these constraints
Incoming
User Traffic
Traffic Forecaster
Workload Profiler
Placement Optimizer
Scheduler
Predicted Traffic
MS
Size
Best Placement
Per MS (Energy,
Latency, Size)
Input from
Application Owner
& External APIs
Microservices
Figure 5: System Components of HuntMS. Modeling thousands of microservices across multiple regions
with latency-aware dependencies creates an enormous search space
that must be navigated efficiently at runtime. For this reason, HuntMS
uses a genetic algorithm (GA) [54, 63], which offers a scalable alternative by evolving a population of candidate placements over
time.Where the paper says it
Where the paper says it
Insight-Driven Pruning
The search space is reduced through key optimizations, including microservice pinning and region filtering. Microservice pinning excludes services that must remain in their base region, such as frontends and databases. p. 6 Microservice Pinning reduces the optimization search space
by excluding microservices that must remain in their base region. First, HuntMS structurally pins microservices that are inherently
immovable due to operational or regulatory constraints, such as
frontends that must remain close to users for responsiveness, and
databases that are bound by data sovereignty laws [25, 50, 65]. Region Filtering further reduces the search space by pruning
suboptimal regions using real-time carbon intensity and pricing
data. Region pruning follows a Pareto-style rule, which retains
regions only if they do not worsen all metrics compared to the
base region.Where the paper says it
Where the paper says it
Evaluation Results
Evaluation on a real-world deployment shows that HuntMS quickly adapts to real-time changes in workload and carbon intensity and reduces carbon emissions by 37.4% and operational cost by 3. p. 0 Evaluation on a real-world deployment
shows that HuntMS quickly adapts to real-time changes in workload and carbon intensity and reduces carbon emissions by 37. 4%
and operational cost by 3. 4%
and operational cost by 3. 6%, on average, compared to a static deployment within a single country, while consistently meeting SLOs. Overall,
HuntMS enables both carbon- and cost-aware microservice placement for latency-sensitive applications in regionally constrained
infrastructures. The specific paper contributions are:
• A characterization of carbon reduction opportunities for microservices within geographically constrained infrastructures,
revealing the insight that selectively offloading early-activated,
non-critical subtrees to greener regions can reduce emissions
without compromising performance (Section 3). Thanks to region filtering and microservice pinning, these placements remain well below the SLO
threshold. In Europe, HuntMS achieves an average end-to-end latency of 93. When load stabilizes, HuntMS
resumes offloading to greener regions such as Stockholm, Paris and
Montreal, restoring carbon and cost benefits. On average, HuntMS
triggers 6–7 placement changes per day, relocating approximately
42% of microservices per change in the European setting and 12%
in the North-America setting. Analysis of Results. Figure 4 compares placement configurations
across four dimensions: carbon emissions, monthly cost, end-toend latency, and SLO violations (Table 3b).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Scalability and Comparison
HuntMS demonstrates scalability, converging in under 2.2 minutes even with 1,000 microservices and all global regions. p. 9 Even with 1,000 microservices and all global regions,
HuntMS converges in under 2. 2 minutes, and takes seconds for
small deployments. Takeaways: Across all metrics, HuntMS offers the most practical
balance of sustainability, performance, and stability, compared to
other baselines. It achieves near-optimal carbon reduction and competitive cost savings while maintaining latency close to the optimal
LP solution.Where the paper says it
Where the paper says it
Conclusion and Future Work
HuntMS enables carbon- and cost-aware microservice deployment for latency-sensitive applications in regionally constrained infrastructures. Future work includes exploring integration with carbon credit markets and developing a dashboard for real-time monitoring to improve usability for SMEs. p. 5 Interfacing with the application owner. HuntMS exposes a
terminal-based configuration interface, with support for a dashboard offering real-time monitoring, parameter updates, and notifications. Such an assumption does not hold for
many national or regional small to medium-sized enterprises (SMEs)
with microservice applications, which represent the real-world majority. In this paper, we present HuntMS, an Adaptive Carbonand Efficiency-aware placement for microservices that considers
carbon, cost, and latency constraints.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive Carbonand Efficiency-aware placement for microservices can dynamically place microservices across geographically constrained regions using a scalable optimization strategy that leverages insight-based search space pruning techniques to reduce carbon emissions by 37.4% and operational cost by 3.6%.
-
The improved AI system can enable carbon- and cost-aware microservice deployment for latency-sensitive applications in regionally limited infrastructures for SMEs, specifically addressing the challenge of deploying microservices within geographically constrained infrastructures while maintaining application performance.
-
HuntMS can solve a constrained multi-objective optimization problem to determine where microservices should be placed across available regions by jointly optimizing carbon emissions and monetary cost while respecting latency constraints (O1: Carbon Efficiency, O2: Cost Efficiency, O3: Latency Compliance).
-
The system can proactively adapt to real-time changes in workload and carbon intensity by using a Traffic Forecaster that runs a Gradient Boosted Decision Tree model every hour for short-term traffic prediction.
-
The improved system can ensure stability under dynamic conditions by using event-driven optimization that re-invokes the Placement Optimizer when triggered by either carbon or workload changes, reacting to forecasted traffic shifts and carbon intensity deviations.
-
HuntMS can reduce the search space dramatically by employing microservice pinning, which pins inherently immovable services like frontends and databases, and ranking microservices based on their activation order to pin later activated ones when end-to-end latency approaches the SLO.
-
The system can prune suboptimal regions using a Pareto-style rule during optimization, retaining only regions that do not worsen all metrics compared to the base region, which is particularly useful for SMEs facing cost constraints where
a region that improves only one metric is discarded if another region strictly improves both carbon and cost over the base region.
-
The system can maintain performance guarantees by focusing on the offloading of microservice subtrees that are activated early in the request execution and do not belong to the critical path, as this
eliminates infeasible or low-impact decisions, reducing optimization overhead while maintaining latency guarantees.
Abstract
Microservices are a dominant architecture in cloud computing, offering scalability and modularity, but also posing complex deployment challenges. As data centers contribute significantly to global carbon emissions, carbon-aware scheduling is a promising mitigation strategy. However, most existing solutions target batch, HPC, or serverless workloads and assume global-scale infrastructure. Such an assumption does not hold for many regional small to medium-sized enterprises (SMEs) with microservice applications, due to geographic, cost, and data sovereignty constraints. In this paper, we present Aceso, an Adaptive Carbon- and Efficiency-aware placement for microservices that considers carbon, cost, and latency constraints. Aceso dynamically places microservices across geographically constrained regions using a scalable optimization with insight-based search space pruning techniques. Evaluation on a real-world deployment shows that Aceso adapts to real-time changes in workload and carbon intensity and reduces carbon emissions by up to 37.4% and operational cost by up to 3.36%, on average across days, compared to a static deployment within a single country, while maintaining the application's latency targets. In this way, Aceso enables carbon- and cost-aware microservice deployment for latency-sensitive applications in regionally limited infrastructures.
Sources
- EcoLearn: Optimizing the Carbon Footprint of Federated Learning
- CarbonEdge: Leveraging Mesoscale Spatial Carbon-Intensity Variations for Low Carbon Edge Computing
- A Survey on Task Scheduling in Carbon-Aware Container Orchestration
Related papers
- Programmable Probabilistic Computer with 1,000,000 p-bits
- BatchGen: An Architecture for Scalable and Efficient Batch Inference
- Seamless Reconfiguration for DAG-BFT (Technical Report)
- ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving
- iScheduler: Reinforcement Learning-Driven Continual Optimization for Large-Scale Resource Investment Problems
- TIRx: A Unified Tile Primitive ML Compiler Across Heterogeneous Backends