SealOS+: A Sealos-based Approach for Adaptive Resource Optimization Under Dynamic Workloads for Securities Trading System

1 st Haojie Jia, 2 nd Zhenhao Li, 3 rd Gen Li, 4 th Minxian Xu*, 5 th Kejiang Ye · 2025

Read SealOS+: A Sealos-based Approach for Adaptive Resource Optimization Under Dynamic Workloads for Securities Trading System by 1 st Haojie Jia, 2 nd Zhenhao Li, 3 rd Gen Li, 4 th Minxian Xu*, 5 th Kejiang Ye (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-23258.txt.

Introduction

Haojie Jia and four co-authors present SealOS Plus for securities trading systems in two thousand twenty-five. Trading systems now run as microservices, which strains resource plans. Delays grow and chips sit idle under old schedulers. Exchanges demand replies in tens of milliseconds.

Their stack adds three parts to a container platform. A deep reinforcement learner schedules resources on the fly. A three level cache speeds common trading steps. A memory based net predicts load ahead of time. In live use at an exchange, chip use rose high. Reply times fell and peak throughput hit a strong rate. Results met the strict demands of the floor.

Catalog record

CREATOR1 st Haojie Jia, 2 nd Zhenhao Li, 3 rd Gen Li, 4 th Minxian Xu*, 5 th Kejiang Ye
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-23258

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SealOS+: A Sealos-based Approach for Adaptive Resource Optimization Under Dynamic Workloads for Securities Trading System 1 st Haojie Jia, 2 nd Zhenhao Li, 3 rd Gen Li, 4 th Minxian Xu*, 5 th Kejiang Ye Abstract As securities trading systems transition to a microservices architecture, optimizing system performance presents challenges such as inefficient resource scheduling and high service response delays. Existing container orchestration platforms lack tailored performance optimization mechanisms for trading scenarios, making it difficult to meet the stringent 50ms response time requirement imposed by exchanges. This paper introduces SealOS+, a Sealos-based performance optimization approach for securities trading, incorporating an adaptive resource scheduling algorithm leveraging deep reinforcement learning, a three-level caching mechanism for trading operations, and a Long Short-Term Memory (LSTM) based load prediction model. Real-world deployment at a securities exchange demonstrates that the optimized system achieves an average CPU utilization of 78%, reduces transaction response time to 105ms, and reaches a peak processing capacity of 15,000 transactions per second, effectively meeting the rigorous performance and reliability demands of securities trading. I Introduction Microservices architecture has emerged as the preferred choice for large-scale distributed systems due to its superior scalability, elasticity, and agility over traditional monolithic architectures [ 1 ] . This shift is particularly advantageous for securities trading systems, which must process massive concurrent transactions with millisecond-level response time (RT). Unlike monolithic systems that struggle with performance bottlenecks—especially during peak trading periods such as market opening and closing—microservices architectures decompose trading systems into independently scalable services (e.g., order processing, matching engine, clearing, and settlement). This decomposition enables better load distribution and resource utilization while improving fault tolerance and maintainability [ 2 ] . Despite these advantages, transitioning securities trading systems to a microservices architecture introduces unique challenges. The first major challenge is extreme latency sensitivity, as modern exchanges impose strict service-level agreements (SLAs) requiring end-to-end order processing to be completed within 50 milliseconds. Any additional delay can lead to missed trading opportunities, financial losses, and a competitive disadvantage. Second, securities trading exhibits highly dynamic workloads, characterized by tidal and bursty loads. Traditional elastic scaling approaches—such as auto-scaling groups in cloud environments—often fail to respond quickly enough to sudden trading surges, leading to resource over-provisioning or performance degradation. Third, microservices introduce significant network overhead due to frequent cross-service and cross-node communication, increasing response latency. Service mesh and API gateway solutions attempt to mitigate these issues but are not specifically optimized for ultra-low-latency trading scenarios, leading to inefficiencies. To address these challenges, we propose a performance optimization scheme called SealOS+ for securities trading systems based on Sealos [ 3 ] , a lightweight cloud operating system that enhances Kubernetes’ scheduling efficiency. Our approach integrates several key techniques to optimize performance under dynamic trading workloads: First, we design an intelligent scheduling algorithm that dynamically adjusts resource allocation based on real-time trading load characteristics. By leveraging deep reinforcement learning, the scheduler learns optimal resource allocation strategies, reducing response latency while maintaining high system throughput [ 4 ] . Second, to mitigate the overhead of frequent inter-service communication, we introduce a caching architecture comprising local memory caching, distributed caching, and persistent storage. This hierarchical caching strategy significantly reduces database query latency and network transmission delays, thereby enhancing overall system responsiveness [ 5 ] . Third, given the bursty nature of trading workloads, we develop a Long Short-Term Memory (LSTM)-based model [ 6 ] that accurately predicts trading volume fluctuations. By proactively adjusting resource allocation ahead of anticipated load spikes, the system achieves more efficient utilization and prevents bottlenecks. The contributions of this paper are: • An adaptive resource scheduling algorithm using deep reinforcement learning to dynamically allocate computing resources in response to changing trading workloads. • A three-level caching architecture to minimize service response time by efficiently managing frequently accessed trading data. • Through real-world deployment at a securities exchange, our optimized system demonstrates substantial performance gains, achieving an average CPU utilization of 78%, reducing transaction response time to 105ms, and supporting peak processing capacities of 15,000 TPS. The rest of the paper is organized as follows: Section II discusses the related work. Section III introduces the system design of our proposed approach and the system model is presented in Section IV. The extensive experiments are shown in Section V, and the conclusions are given in Section VI. II Related Work Existing research on container orchestration and resource scheduling can be broadly classified into three categories: (1) scheduling and load balancing techniques, (2) container orchestration for performance optimization, and (3) machine learning-based resource management. While these studies provide valuable insights into general cloud computing environments, they do not fully address the unique challenges posed by bursty trading workloads, ultra-low-latency constraints, and real-time adaptive resource allocation. II-A Scheduling and Load Balancing Techniques Several studies have explored scheduling strategies to improve resource utilization and load balancing in cloud and containerized environments. Senthil et al. [ 7 ] introduced a hybrid Genetic Algorithm-Ant Colony Optimization (GA-ACO) scheme for energy-efficient load balancing, but it does not account for the unpredictable surges in trading workloads. Nguyen and Kim [ 8 ] proposed a Kubernetes-based load balancing approach to enhance scalability; however, their method does not incorporate business-specific constraints, such as trading latency requirements. Jeong and Jeong [ 9 ] developed ARAScaler, an adaptive resource scaling system, but it fails to meet the millisecond-level response time needed for financial trading systems. Existing scheduling approaches either focus on general cloud applications or prioritize energy efficiency over real-time responsiveness. Unlike these works, our Sealos-based approach employs deep reinforcement learning (DRL) to dynamically adjust scheduling policies based on real-time trading workload patterns, ensuring minimal response latency and efficient resource utilization. II-B Container Orchestration for Performance Optimization Several researchers have investigated performance optimization in containerized environments. Reddy et al. [ 10 ] analyzed Kubernetes performance across OpenStack, virtual machines, and bare-metal deployments, identifying significant performance variations but not proposing optimization strategies. Marchese and Tomarchio [ 11 ] designed a network SLO-based container orchestration framework, though its centralized architecture introduces bottlenecks during high-frequency trading peaks. Carrión [ 12 ] and Casalicchio [ 13 ] surveyed Kubernetes scheduling challenges, emphasizing the inadequacies of existing strategies in handling bursty workloads. Prior research lacks end-to-end performance optimization strategies tailored to financial trading systems. In contrast, our work introduces a three-level caching mechanism (local memory, distributed cache, and persistent storage) to reduce inter-service communication latency, improving responsiveness under high transaction loads. II-C Machine Learning-Based Resource Management Recent advancements have applied machine learning to container scheduling and resource allocation. Xu and Buyya [ 14 ] proposed a Brownout-based adaptive resource management framework for cloud environments, focusing on general cloud applications rather than trading workloads. Zhong et al. [ 15 ] categorized machine learning-based container orchestration techniques, identifying high inference overhead as a major limitation. Jian et al. [ 16 ] developed DRS, a deep reinforcement learning-based Kubernetes scheduler that dynamically optimizes container placement and resource allocation. However, it primarily targets load balancing without addressing trading-specific constraints such as sub-50ms latency. Li et al. [ 17 ] proposed a reinforcement learning-based rate limiter for e-commerce services, which does not meet the high-frequency trading requirements. Existing machine learning-based approaches either lack real-time inference capabilities or focus on general cloud applications rather than

Read the complete text