Multi-Dimensional Autoscaling of Stream Processing Services on Edge Devices
Summary: arXiv:2510.06882v2 Announce Type: replace-cross
Edge devices, characterized by their limited computational and memory resources, often struggle to meet the demands of stream processing services. This limitation can lead to scenarios where these services fail to satisfy their performance requirements, commonly referred to as Service Level Objectives (SLOs). While traditional autoscaling techniques primarily focus on resource scaling, it is essential to explore alternative strategies that cater specifically to the unique challenges faced by edge devices.
Introducing MUDAP
To address the resource constraints and enhance the performance of stream processing services, we present the Multi-dimensional Autoscaling Platform (MUDAP). This innovative framework supports fine-grained vertical scaling across both service- and resource-level dimensions, allowing for a more tailored approach to autoscaling.
Key Features of MUDAP
MUDAP is designed to optimize the execution of services on edge devices by implementing service-specific scaling strategies. The platform allows for adjustments based on various parameters, including:
- Data quality
- Model size
- Processing speed
- Resource allocation
Scaling Agent: RASK
To further enhance the autoscaling capabilities of MUDAP, we introduce a scaling agent known as Regression Analysis of Structural Knowledge (RASK). The RASK agent employs advanced regression techniques to efficiently explore the solution space and develop a continuous regression model of the processing environment. This model enables RASK to infer optimal scaling actions, ensuring that services can dynamically adjust in response to varying workloads.
Performance Comparison
We conducted a comparative analysis of our approach against two established autoscalers: the Kubernetes Vertical Pod Autoscaler (VPA) and a reinforcement learning-based agent. The experiments involved scaling up to 9 services on a single edge device, with a focus on measuring the effectiveness of each approach in maintaining SLOs under different load conditions.
Results
The results of our experiments demonstrated that the RASK agent was capable of inferring an accurate regression model within just 20 iterations, corresponding to approximately 200 seconds of processing time. By progressively incorporating additional elasticity dimensions into the scaling process, RASK achieved a remarkable reduction in SLO violations, sustaining the highest request load with 28% fewer violations compared to the baseline methodologies.
Conclusion
In conclusion, the Multi-dimensional Autoscaling Platform (MUDAP) offers a promising solution to the challenges of resource management in edge devices. By leveraging the capabilities of the RASK scaling agent, MUDAP can optimize service performance while adhering to stringent SLOs. The findings from our research underline the importance of innovative autoscaling strategies tailored to the unique constraints of edge computing environments.