Scheduling across cloud and edge systems must handle job heterogeneity, dynamic arrivals, and coupled resource constraints. To address the above issues, we present GT-MARL, a reinforcement learning framework with graph and Transformer enhancements for joint task selection and resource allocation across multiple clusters. At each decision epoch, GT-MARL encodes a heterogeneous graph over tasks and resources with Hierarchical Attention Network (HAN) to capture structural dependencies, applies a temporal Transformer to model backlog evolution and delayed interactions, and produces factorized decisions for each cluster through a Transformer-guided task selection head and discrete resource allocation with a Multi-Layer Perceptron (MLP) head. Training follows centralized training with decentralized execution (CTDE) using a centralized critic and masked action spaces. Experiments on both synthetic and trace-replay workloads show that GT-MARL consistently achieves the best p95 job completion time (JCT), while keeping mean JCT competitive with representative heuristic and learning baselines. Specifically, GT-MARL obtains 35.56 mean JCT and 70.76 p95 JCT on the synthetic workload, and 144.21 mean JCT and 264.47 p95 JCT on the trace-replay workload. On the synthetic workload, this corresponds to a 26.3% p95 JCT reduction relative to the best mean baseline with only a 3.0% increase in mean JCT. Additional queueing and service decomposition across workload intensities indicates that the tail-latency gain mainly comes from alleviating queue accumulation rather than shortening the inherent service time of tasks. GT-MARL provides a favorable tradeoff between mean latency and tail latency for scheduling under SLO constraints across cloud and edge systems.

During software maintenance and evolution, developers spend more than half of their time on code comprehension activities. In order to understand an unfamiliar code base, they would naturally ask different types of questions related to code snippets and try to find the answers. In this paper, we conduct an initial work to explore the possibility of automatic question generation for program comprehension. We construct a large-scale data set containing pairs of source code and questions that are automatically transformed from inline comments based on dependency analysis and semantic role labeling. We also build a comprehensive taxonomy of question types so as to generate questions concerning different aspects of code snippets, such as purpose, implementation details and so on. Then, we propose a deep learning-based prototype CodeQG to automatically generates multiple types of questions for code snippets. We evaluate CodeQG by using both typical performance metrics and manual evaluation. The results show that (1) we can achieve a value of 42.02 on BLEU4 and 60.81 on ROUGE-L for the generated questions; (2) overall, the questions are very correct in grammatical, semantic and format; (3) the questions are related to the corresponding code snippet and are helpful for developers in source code comprehension activities. Our work gives insights into automatically generating multiple types of questions for code comprehension. We expect this exploration will improve the applicability and generality of machine code comprehension.

With the continuous development in the field of machine learning, there is an increasing demand for the cloud-based machine learning inference services, which are latency-sensitive tasks, such as the service requests from the Internet of Things (IoT) devices. These inference services are generally accompanied by fluctuations and uncertainties, so they often require vastly varied numbers of servers at different time spots. As a result, how to dynamically and rationally schedule cloud servers for inference services has become an important issue. Alibaba Cloud currently provides its serverless instances called elastic container instance (ECI), and due to the advantages of pay-as-you-go billing and second-level elasticity, they are well-suited for handling bursty or fluctuating workloads. At the same time, Alibaba Cloud’s subscription-based elastic compute service (ECS) instances can be used for steady-state workloads. Our objective is to dynamically combine these two types of instances to deal with inference service requests. In this article, we propose a deterministic online algorithm that can rationally schedule these two types of instances to optimize costs without requiring knowledge of future workloads. We prove that the proposed online algorithm achieves a competitive ratio of no more than 2 compared to the optimal offline algorithm. Through simulation experiments, we demonstrate that our algorithm outperforms three benchmarks, which are all-reserved algorithm, all-on-demand algorithm, and traditional online algorithms that only use ECS instances. Our algorithm exhibits superiority across various workloads and can significantly reduces costs in most cases.

At present, the microservice architecture is widely used in modern software development for its flexibility and scalability. However, the huge data scale and complex invocation relationships between services make the root cause localization of faults in microservice systems extremely difficult. Some of the current root cause localization methods based on metrics data use history fault data to effectively improve the localization results, but there are challenges such as not representing history faults effectively and limiting the localization results to repetitive faults that have occurred in history. In this paper, we propose an automatic root cause localization framework MicroHFRCL based on the history fault library to address the above issues. MicroHFRCL constructs an instance causal graph based on metric data for causal analysis. The instance causal graph is weighted by encoding the anomalous subgraph and calculating the similarity of history faults. The PageRank algorithm is used to locate the root cause of faults. Among them, MicroHFRCL learns the structure and feature information of fault anomalous subgraphs through GCN and Transformer models, achieving effective representation of history faults and fast calculation of similarity in history fault codes, improving the efficiency of repetitive fault localization, and effectively solving the problem of the limitation of using history faults for root cause localization results. We implemented MicroHFRCL and tested it on the fault dataset collected by a benchmark microservice test system. Compared with the latest baseline models, MicroHFRCL has significantly improved the localization accuracy, and can also achieve good results in the case of small-scale history faults.

The recent trend of "everything as a service", is fostering the Internet of services (IoS) and will promote the emergence of service operation. Service operation provides the business and technical base for advanced business models. Service monitoring is a crucial issue for the guaranteed service delivery in service operation. However, most service monitoring approaches are specific and focus on IT level. The challenge of how to monitor diverse business aspect of service simply and flexibly needs to be overcome. In this paper, we present a model-driven approach for service monitoring from business perspective. In the approach, a business-oriented service monitoring metamodel is put forward to define various monitoring models on demand. The model can flexibly specify the monitored information in both business level and IT level and the monitoring process. Also, a service monitor is implemented through model-driven way which brings the scalability to its implementation.

Web services are rapidly emerging as a popular standard technology for sharing data and functionality among heterogeneous systems. Service providers and consumers are loosely coupled and distributed across the network, either within an organization or across organizational boundaries, and therefore, performance becomes a major concern in such a distributed environment. Furthermore, XML is widely used as message format for service providers and consumers in Web services environment. XML message packaging and parsing brings extra overhead to both ends. Web services response latency, as well as throughput, is becoming a bottleneck problem. In this paper, We propose a consistency-preserving mechanism for Web services response caching, which reduces the volume of data transmitted without semantic interpretation of service requests or responses, and accelerates the services response finally. It achieves this reduction through the use of cryptographic hashing to detect similarities with previous results. Experiments with an initial prototype called SigsitAcclerator indicate that our mechanism can lead to significant performance improvement over more straightforward techniques.