Cloud requests and infrastructures are gradually showing a trend of geographical distribution, and the scale is getting bigger and bigger. Due to a lack of information about the network, resources, and other tenants, tenant-made resource region decisions will incur more unnecessary costs than those made by the cloud provider. For such a complex scenario, public cloud providers even put stringent efficiency requirements, requiring decision-making within 1 s for 100 K requests. In summary, there are three significant challenges for task placement in large-scale geographically distributed clouds: complex topology, large scale, and lagging information. Existing studies either focus on intra-regional task placement only or ignore the complex request topology with large-scale and strict decision delay constraints, and most of them need to consider the impact of information lag on decision-making. To address these challenges, we design a large-scale task placement framework called Lattice. Despite the tough challenges posed on a large scale, we find that there are many similar relationships among complex topologies. Therefore, combining offline characterization and online placement, we design a data structure and algorithm based on pattern matching topology compression and interval retrieval, a multi-dimensional parallel placement architecture, and a competition-aware state representation strategy to achieve efficient and high-quality task deployment. Simulation results based on real-world datasets show that our algorithm can achieve an efficiency of 1.74x-15.01x that of existing methods while maintaining similar decision quality. We have open-sourced our algorithm, experimental environment, and experimental data.

In a cluster, a state manager is often used to maintain the states of all hosts and provide the states to schedulers for task scheduling. With the evolution of cloud computing, cluster scales have been expanding. In a large-scale cluster, the state manager, constrained by its limited processing capacity, tends to synchronize with hosts periodically. As a result, resources released by tasks remain marked as occupied in the state manager's view and cannot be reallocated until the next synchronization, which is harmful to the utilization of cluster resources. We refer to these unsynchronized but released resources as Shadow Resources (SR). According to our theoretical analysis, for a cluster with 4k hosts that appears fully loaded in the state manager's view, the shadow resources can account for up to 6.22%. Most existing studies overlook shadow resources, while a few attempt to mitigate their impact by introducing host task queues. However, this approach relies on idealized task runtime prediction and increases system complexity, making it challenging to implement in practice. To overcome the challenges of collecting fleeting shadow resources and utilizing synchronized resources effectively, we propose Shadow Resource Management Architecture (SRMA), which can be embedded into a large-scale cluster under global management. SRMA logically partitions the cluster and collects the shadow resources within each partition through frequent state synchronization. An external SR load balancer, employing the Max-Resource partition First (MRF) algorithm, assigns tasks to SR schedulers, which allocate the collected shadow resources for the tasks. Experiments show that SRMA effectively utilizes shadow resources, achieving near-ideal resource utilization without synchronization delays. It also outperforms other related architectures in task latency and throughput time and provides a benefit equivalent to an increase of 8.5% in the resources of the original cluster.

Resource wastage is common in GPU clusters running deep learning training (DLT) tasks. Studies show that improper allocation of CPU resources is one of the main factors contributing to this waste. Existing methods can predict the task durations under different GPU and CPU joint allocation schemes in advance and then guide resource allocation scheduling to reduce resource wastage. However, the existing prediction methods are challenging in meeting the prediction accuracy and efficiency requirements of the complex combination of GPUs and CPUs. Meanwhile, the existing resource allocation scheduling methods only consider individual task optimization and cannot achieve global optimal allocation. To address these weaknesses, we propose CoreTuner, a predicting and scheduling framework for the joint allocation of CPU and GPU. Firstly, we define a CPU influence model and an Optimal CPU Core Count (OCCC) metric that quantifies how increasing CPU cores beyond a certain threshold can degrade performance. Secondly, based on this model and the OCCC metric, we propose a multi-schemes duration prediction algorithm that combines sampling and extrapolation, achieving accurate multi-schemes prediction with only one sampling, balancing prediction efficiency and accuracy. Finally, we designed a cluster-level dynamic optimal scheduling algorithm. This algorithm prioritizes resource allocation to the tasks that are most sensitive to resources, thereby reducing the overall execution time of tasks across the cluster. Experimental results show that CoreTuner can significantly improve the cluster’s performance: the task makespan, average turnaround time, and GPU utilization increase by up to 51.3%, 75.3%, and 56.54%, respectively, compared to the existing algorithms.

RoCE (RDMA over Converged Ethernet) performs well in a lossless network. However, its performance drops dramatically when packet loss occurs. This limitation stems from the shortcomings of traditional solutions in packet loss recovery, which are explicitly reflected in the two stages of packet loss detection and packet loss retransmission. For packet loss detection, they either rely on implicit detection, which leads to misjudgment or use explicit notification methods that cannot effectively detect multiple lost packets. For packet loss retransmission, they either adopt the Go-back-N mechanism, which is inefficient or implement selective retransmission that still fails to retransmit all missing segments accurately. To address these issues, we propose SNACK, a novel packet recovery mechanism that combines explicit notification and selective NACK(negative acknowledgment) retransmission. SNACK brings two key improvements. First, it precisely detects lost packets through dynamic window monitoring. Second, it adapts to retransmission using SNACK frames, which can selectively request missing packets. Our NS-3 simulation results show that SNACK achieves strong performance under packet loss. When the loss rate is 1%, SNACK keeps the throughput within 10% of the lossless case. In multiflow scenarios, it reduces the average FCT (Flow Completion Time) by 24.8% compared to existing methods. These results demonstrate that SNACK provides efficient loss recovery in various network conditions.

Current IaaS providers have deployed data centers worldwide, with resources continually increasing. Meanwhile, there is a rising trend in the concurrency of user requests and the diversity of user request types. To achieve better resource allocation, various complex scheduling architectures have been proposed. However, due to the challenges associated with real-world experiments, simulation systems are needed to build exper-imental environments for related research. As existing systems do not perform well enough, we construct LGDCloudSim. It is designed with full consideration of the characteristics of the large-scale geographically distributed cloud data center scenarios. To support large-scale simulations, we propose state management optimization and operation process optimization methods. Exper-iments show that LGDCloudSim can simulate up to 5 X 108hosts and 107request concurrency. It also supports diverse scheduling architectures and different request types.

Monolithic architectures are becoming increasingly difficult to cope with complex applications, and microservice architectures, which offer flexibility and logical independence in development and maintenance, are the new choice for companies and developers. Migrating a legacy monolithic architecture application to a microservice architecture rather than building it from scratch is considered an easy way to use it. To ensure that the migrated microservice applications can take advantage of their benefits, we need to propose a reasonable and effective microservice extraction method. Considering the single responsibility principle in the microservice design principle, most existing microservice extraction methods only pursue the high logical independence of the extraction results and pay little attention to whether the extraction results have good performance. Applications need to perform well, and studies have shown that poor microservice extraction schemes can negatively impact the performance of the migrated application. As a result, when extracting, we should also consider the performance of the results. A few studies consider the performance of extraction results, but only in terms of a few factors affecting performance, such as network overhead, rather than considering all factors affecting performance comprehensively, which leads to an inaccurate evaluation of performance. Therefore, oriented toward the most widely used managed languages today, we propose an effective Microservice Extraction method based on a Comprehensive Evaluation of logical independence and performance (MECE). Firstly, we propose a workflow-based approach to evaluate the performance of microservice extraction results by considering multiple influencing factors, focusing on the management cost ignored in existing studies, and designing an effective management cost evaluation model. After that, we propose a meta-heuristic search-based algorithm to obtain feasible microservice extraction results. In experiments based on actual deployments, the extraction results of the MECE method obtained a performance improvement of up to 46.15% without significant loss of logical independence compared to existing methods, which verifies the effectiveness of the method.

Community search is a fundamental problem in social network analysis. In this paper, we propose a novel problem, called influential maximum connected community (IMCC) search, which leverages the k-ECC model to measure the cohesiveness of subgraphs and tries to find the influential community for a set of query vertices.