Software systems often expose a large number of configurable parameters to satisfy diverse application requirements and deployment scenarios. Given the intricate dependencies between parameters, manually finding a well-performing configuration is a daunting task even for experienced operators. Most existing automatic tuning approaches treat the problem as a single-objective search, leaving critical concerns such as energy consumption and reliability as afterthoughts. Although recent studies have explored multi-objective configuration tuning, they still face several challenges, including handling conflicting objectives, balancing search effectiveness and efficiency, and adapting to heterogeneous configuration spaces across different software systems. To address these issues, we propose PFTuner, an efficient and effective multi-objective configuration tuning framework adaptive to diverse software systems. PFTuner consists of three collaborative modules, namely Configuration Generator, Configuration Evaluator, and Sample Collector, which operate iteratively to continuously improve configuration quality. In particular, we design a novel multi-objective optimization algorithm that effectively models heterogeneous configuration spaces and improves the balance between optimization quality and search efficiency. We evaluate PFTuner on eight software-workload scenarios deployed on a local cluster and compare it with several representative state-of-the-art baselines. Experimental results show that PFTuner consistently achieves higher-quality Pareto fronts and better search efficiency, while also demonstrating strong adaptability across different software systems and workloads.

Accurate and robust anomaly detection is critical for microservice system reliability. Recent multimodal approaches have improved detection comprehensiveness by integrating metrics, logs, and traces. However, they often overlook intra-modal uncertainty from noise, ambiguity, or missing data, and inter-modal uncertainty arising from varying predictive capabilities across modalities. Additionally, extensive labeling of multimodal data remains costly. To address these limitations, we propose MUAD, an uncertainty-aware multimodal anomaly detection framework with active learning. MUAD employs a Graph-based Probabilistic Encoder (GPE) to model intra-modal uncertainty through probabilistic representations, and a Confidence-aware Fusion Mechanism (CFM) to dynamically weight modalities based on their prediction confidence. Furthermore, an active learning paradigm iteratively refines the model using high-confidence pseudo-labels and informative samples, maintaining performance under label-deficient conditions. Experiments on three benchmark datasets demonstrate MUAD achieves 98.10% average F1-score, outperforming state-of-the-art methods by up to 7.87%. Results also confirm its robustness under low-quality data and limited labels.

Due to the dynamic workloads and complex interactions among different microservices, it is difficult for web service providers to elastically allocate cloud resources to each interior microservice to minimize the total resource usage within the service-level objective (SLO) constraint. Since proactive resource management relies on historical data and complex machine learning models to forecast future workload patterns and resource demands, recent efforts have shifted towards reactive resource management. However, previous studies still suffer from serious challenges including: (1)accurately locating and explaining performance bottlenecks, (2)ensuring SLO compliance during online configuration exploration, (3)identifying and configuring the performance-critical parameters inherent in bottleneck microservices. To overcome these challenges, this paper introduces iSafeRM, an interpretable and safety-aware resource management framework tailored for microservice-based systems. Upon detecting an SLO violation, iSafeRM automatically locates the bottleneck microservices and performs a safety-aware online configuration process of both computing resources and performance-critical parameters. Experimental results conducted on our laboratory Docker cluster show that compared with four representative baselines, iSafeRM can respectively reduce resource usage by an average of 7.5% to 19.1%, 9.3% to 23.0% and 18.7% to 31.0% across three different applications, while maintaining the lowest probability of SLO violations. We also demonstrate iSafeRM's effectiveness under dynamic workloads.

Multivariate time series anomaly detection (MTS-AD) is of great significance in various modern industrial applications and IT systems. Recently, some unsupervised deep models have been developed for MTS-AD. However, these methods often struggle to handle the complex temporal patterns and inevitable noise in MTS data, resulting in limited performance. To overcome these challenges, we propose DeAnomaly, a novel anomaly detection framework based on time series decomposition. Specifically, DeAnomaly employs a two-phase training paradigm, consisting of structural pattern elimination and anomaly detection on remainders. The structural pattern elimination phase learns normal trend and seasonal components through spatial relationship modeling and time-frequency analysis, which are subsequently removed from the original time series to overcome the limitation of complex temporal patterns. The anomaly detection phase utilizes the robust characteristics of noise with denoising diffusion models to identify and distinguish between noise and actual anomalies. Since anomalies and small random fluctuations are mainly retained in the remainders, anomalies will be more clearly exposed. In this way, DeAnomaly can detect anomalies more accurately and robustly. We conduct extensive experiments on four real-world datasets and 13 baselines, experimental results demonstrate that DeAnomaly outperforms these state-of-the-arts.

The performance of modern DBMSs such as MySQL and PostgreSQL heavily depends on the configuration of performance-critical knobs. Manual tuning these knobs is laborious and inefficient due to the complex and high-dimensional nature of the configuration space. Among the automated tuning methods, reinforcement learning (RL)-based methods have recently sought to improve the DBMS knobs tuning process from several different perspectives. However, they still encounter challenges with slow convergence speed during offline training. In this paper, we mainly focus on how to leverage the valuable tuning hints contained in various textual documents such as DBMS manuals and web forums to improve the offline training of RL-based methods. To this end, we propose an efficient DBMS knobs tuning framework named DemoTuner via a novel LLM-assisted demonstration reinforcement learning method. Specifically, to comprehensively and accurately mine tuning hints from documents, we design a structured chain of thought prompt to employ LLMs to conduct a condition-aware tuning hints extraction task. To effectively integrate the mined tuning hints into RL agent training, we propose a hint-aware demonstration reinforcement learning algorithm HA-DDPGfD in DemoTuner. As far as we know, DemoTuner is the first work to introduce the demonstration reinforcement learning algorithm for DBMS knobs tuning. Experimental evaluations conducted on MySQL and PostgreSQL across various workloads demonstrate that DemoTuner achieves performance gains of up to 44.01

Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks. However, they remain exposed to jailbreak attacks, eliciting harmful responses. The nested scenario strategy has been increasingly adopted across various methods, demonstrating immense potential. Nevertheless, these methods are easily detectable due to their prominent malicious intentions. In this work, we are the first to find and systematically verify that LLMs' alignment defenses are not sensitive to nested scenarios, where these scenarios are highly semantically relevant to the queries and incorporate targeted toxic knowledge. This is a crucial yet insufficiently explored direction. Based on this, we propose RTS-Attack (Semantically Relevant Nested Scenarios with Targeted Toxic Knowledge), an adaptive and automated framework to examine LLMs' alignment. By building scenarios highly relevant to the queries and integrating targeted toxic knowledge, RTS-Attack bypasses the alignment defenses of LLMs. Moreover, the jailbreak prompts generated by RTS-Attack are free from harmful queries, leading to outstanding concealment. Extensive experiments demonstrate that RTS-Attack exhibits superior performance in both efficiency and universality compared to the baselines across diverse advanced LLMs, including GPT-4o, Llama3-70b, and Gemini-pro. Our complete code is available at https://github.com/nercode/Work. WARNING: THIS PAPER CONTAINS POTENTIALLY HARMFUL CONTENT.

To ensure high-performance processing capabilities across diverse application scenarios, Big Data frameworks such as Spark and Flink usually provide a number of performance-related parameters to configure. Considering the computation scale and the characteristic of repeated executions of typical recurring Big Data processing jobs, how to automatically tune parameters for performance optimization has emerged as a hot research topic in both academic and industry. With the advantages in interpretability and generalization ability, causal inference-based methods recently prove their advancement over conventional search-based and machine learning-based methods. However, the complexity of Big Data frameworks, the time-varying input dataset size of a recurring job and the limitation of a single causal structure learning algorithm together prevent these methods from practical application. Therefore, in this paper, we design and implement CausalConf, a datasize-aware configuration auto-tuning approach for recurring Big Data processing jobs via adaptive causal structure learning. Specifically, the offline training phase is responsible for training multiple datasize-aware causal structure models with different causal structure learning algorithms, while the online tuning phase is responsible for recommending the next promising configuration in an iterative manner via the Multi-Armed Bandit-based optimal intervention set selection as well as the novel datasize-aware causal Bayesian optimization. To evaluate the performance of CausalConf, a series of experiments are conducted on our local Spark cluster with 9 different previously unknown target applications from HiBench. Experimental results show that the performance speed ratio achieved by CausalConf compared to the four recent and representative baselines can respectively reach 1.45x, 1.31x, 1.26x and 1.54x on average and up to 2.53x, 1.55x, 1.57x, 2.18x. Besides, the average total online tuning cost of CausalConf is reduced by 8.85%, 14.26%, 18.58%, and 14.29%, respectively.

Considering the operational efficiency and cost-effectiveness, big data pipelines are increasingly deployed on the cloud. Since streaming jobs often recur at specific intervals, selecting appropriate resource configuration and software parameters for these recurring jobs can significantly reduce resource costs. However, the high-dimensional configuration space and the time and cost-consuming configuration evaluation together makes this joint optimization problem quite difficult if not impossible. To address this challenging problem, this paper introduces a framework called COTuner which is able jointly auto-tune the resource configuration and software parameters for recurring streaming jobs deployed on the cloud. Specifically, COTuner first introduces a pseudo-point mechanism into the BO_dropout algorithm to address the formulated high-dimensional black-box optimization problem while improving the convergence rate. Besides, COTuner is also able to be quick adaptive to different scenarios via the ability to adaptively adjusting its hyperparameters. We evaluated the effectiveness and efficiency of COTuner with three different stream processing jobs from Hibench under a local streaming pipeline consisting of Kafka and Flink. Experimental results show that COTuner can achieve better resource costs than all the other four baseline methods under the same performance constraint. In addition, we also leverage ablation experiments to analyze and illustrate the function of each proposed techniques in COTuner.

Serverless computing, also known as Function-as-a-Service (FaaS), is a significant development trend in modern software system architecture. The workflow composition of multiple short-lived functions has emerged as a prominent pattern in FaaS, exposing a considerable resources configuration challenge compared to individual independent serverless functions. This challenge unfolds in two ways. Firstly, workflows frequently encounter dynamic and concurrent user workloads, increasing the risk of QoS violations. Secondly, the performance of a function can be affected by the resource reprovision of other functions within the workflow. With the popularity of the mode of concurrent processing in one single instance, concurrency limit as a critical configuration parameter imposes restrictions on the capacity of requests per instance. In this study, we present FaaSConf, a QoS-aware hybrid resource configuration approach that uses multi-agent reinforcement learning (MARL) to configure hybrid resources, including hardware resources and concurrency, thereby ensuring end-to-end QoS while minimizing resource costs. To enhance decision-making, we employ an attention technique in MARL to capture the complex performance dependencies between functions. We further propose a safe exploration strategy to mitigate QoS violations, resulting in a safer and efficient configuration exploration. The experimental results demonstrate that FaaSConf outperforms state-of-the-art approaches significantly. On average, it achieves a 26.5% cost reduction while exhibiting robustness to dynamic load changes.

Big data frameworks usually provide a large numberof performance-related parameters. Online auto-tuning these pa-rameters based on deep reinforcement learning (DRL) to achieve abetter performance has shown their advantages over search-basedand machine learning-based approaches. Unfortunately, the timecost during the online tuning phase of conventional DRL-basedmethods is still heavy, especially for Big Data applications. There-fore, in this paper, we propose DeepCAT+, a low-cost and trans-ferrable deep reinforcement learning-based approach to achieveonline configuration auto-tuning for Big Data frameworks. Toreduce the total online tuning cost and increase the adaptability: 1)DeepCAT+utilizes the TD3 algorithm instead of DDPG to alleviatevalue overestimation; 2) DeepCAT+modifies the conventional ex-perience replay to fully utilize the rare but valuable transitions viaa novel reward-driven prioritized experience replay mechanism;3) DeepCAT+designs a Twin-Q Optimizer to estimate the exe-cution time of each action without the costly configuration eval-uation and optimize the sub-optimal ones to achieve a low-costexploration-exploitation tradeoff; 4) Furthermore, DeepCAT+also implements an Online Continual Learner module based onProgressive Neural Networks to transfer knowledge from historicaltuning experiences. Experimental results based on a lab Spark clus-ter with HiBench benchmark applications show that DeepCAT+is able to speed up the best execution time by a factor of 1.49x,1.63xand 1.65xon average respectively over the baselines, whileconsuming up to 50.08%, 53.39% and 70.79% less total tuningtime. In addition, DeepCAT+also has a strong adaptability to thetime-varying environment of Big Data frameworks

Hyperparameters optimization (HPO) is vital for machine learning models. Besides model accuracy, other tuning intentions such as model training time and energy consumption are also worthy of attention from data analytic service providers. Therefore, it is essential to take both model hyperparameters and system parameters into consideration to execute cross-layer multi-objective hyperparameter auto-tuning. Toward this challenging target, we propose HyperTuner in this paper which leverages a well-designed ADUMBO algorithm to find the Pareto-optimal configuration set. Compared with vanilla Bayesian optimization-based methods, ADUMBO selects the most promising configuration from the generated Pareto candidate set during each iteration via maximizing a novel adaptive uncertainty metric. We evaluate HyperTuner on our local distributed TensorFlow cluster, and experimental results show that it is always able to find a better Pareto configuration front superior in both convergence and diversity compared with the other four baseline algorithms. Besides, experiments with different training datasets, different optimization objectives, and different machine learning platforms verify that HyperTuner can well adapt to various data analytic service scenarios.

Energy-efficiency now severely restricts the sustainable operation and development of big data services. In this paper, we propose a bi-objective configuration parameters auto-tuning method EFTuner towards energy-efficient big data processing. Following the sampling-modeling-searching workflow, EFTuner first leverages the Latin Hypercube Sampling to collect configuration sample data under multiple dataset input sizes and then separately build a datasize-aware prediction model for performance and energy consumption with Stochastic Gradient Boosted Regression Tree. Besides, to avoid meaningless variation in the evolutionary process of original NSGA-II, EFTuner also explores the Pareto-optimal configurations with a novel parameter importance-based mutation operation. Experiments conducted on a local 3-node Spark cluster with three different applications verify the advantages of EFTuner over the baselines.

Big data processing frameworks such as Spark usually provide a large number of performance-related configuration parameters, how to auto-tune these parameters for a better performance has been a hot issue in academia as well as industry for years. Through delicately tradeoff between exploration and exploitation, Bayesian Optimization (BO) is currently the most appealing algorithm to achieve configuration auto-tuning. However, considering the tuning cost constraint in practice, there are three critical limitations preventing conventional BO-based approaches from being directly applied into auto-tuning cluster-based big data frameworks. In this paper, we propose a cost-efficient configuration auto-tuning approach named TurBO for big data frameworks based on two enhancements of vanilla BO:1) To reduce the essential iteration times, TurBO integrates a well-designed adaptive pseudo point mechanism with BO; 2) To avoid the time-consuming practical evaluation of sub-optimal configurations as possible, TurBO leverages the proposed CASampling method to intelligently tackle with these sub-optimal configurations based on ensemble learning with historical tuning experiences. To evaluate the performance of TurBO, we conducted a series of experiments on a local Spark cluster with 9 different HiBench benchmark applications. Overall, compared with 3 representative BO-based baseline approaches OpenTuner, Bliss and ResTune, TurBO is able to speedup the tuning procedures respectively by 2.24×, 2.29× and 1.97× on average. Besides, TurBO can always achieve a positive cumulative performance gain under the simulated dynamic workload scenario, which means TurBO is indeed appropriate for workload changes of big data applications.

To support different application scenarios, big data frameworks usually provide a large number of performance-related configuration parameters. Online auto-tuning these parameters based on deep reinforcement learning to achieve a better performance has shown their advantages over search-based and machine learning-based approaches. Unfortunately, the time consumption during the online tuning phase of conventional DRL-based methods is still heavy, especially for big data applications. Therefore, in this paper, we propose DeepCAT, a cost-efficient deep reinforcement learning-based approach to achieve online configuration auto-tuning for big data frameworks. To reduce the total online tuning cost: 1) DeepCAT utilizes the TD3 algorithm instead of DDPG to alleviate value overestimation; 2) DeepCAT modifies the conventional experience replay to fully utilize the rare but valuable transitions via a novel reward-driven prioritized experience replay mechanism; 3) DeepCAT designs a Twin-Q Optimizer to estimate the execution time of each action without the costly configuration evaluation and optimize the sub-optimal ones to achieve a low-cost exploration-exploitation trade off. Experimental results based on a local 3-node Spark cluster and HiBench benchmark applications show that DeepCAT is able to speed up the best execution time by a factor of 1.45 × and 1.65 × on average respectively over CDBTune and OtterTune, while consuming up to 50.08% and 53.39% less total tuning time.

To support different application scenarios, graph databases (GDBs) usually provide a large number of performance-related parameters for developers. Since manually configuring is both time-consuming and cost-intensive, automatically tuning configurations parameters to achieve a better performance has been an urgent need. Besides, considering various graph management requirements, GDBs begin to utilize the modular architecture to interoperate with a wide range of storage and index backends. Due to the complicated interactions among different modules, sequentially tuning each software with previous solutions may fall into a local optimal and it is necessary to jointly autotune the cross-module configuration parameters. Toward this challenging target, we propose JointConf-a new black-box approach of jointly autotuning configuration parameters for modularized GDBs. To address the formulated high-dimensional black-box optimization problem, JointConf utilizes the recently proposed BO_dropout algorithm. Inspired by the dropout algorithm in neural networks, BO_dropout explores efficient dimension dropout to achieve a high-dimensional Bayesian optimization. We evaluate the effectiveness of JointConf on a local distributed JanusGraph cluster with three different graph query benchmark applications and experimental results show its advantages over the four baseline search-based approaches. The necessity of jointly tuning for modularized GDBs is also verified in our experiments.

To support various application scenarios, big data processing frameworks (BDPFs) such as Spark usually provide users with a large number of performance-critical configuration parameters. Since manually configuring is both labor-intensive and time-consuming, automatically tuning configurations parameters for BDPFs to achieve better performance has been an urgent need. To simultaneously address the corresponding challenges such as high dimensional configuration space, we propose ATConf -a new black-box approach of automatically tuning the internal and external configuration parameters for BDPFs. Experimental results based on our local distributed Spark cluster show that the best execution time achieved by ATConf is as much as 46.52% less than the default configuration. Besides, compared with the four baselines, ATConf is able to further reduce the relative execution time over default by at least 4.10% under the same constraint of observation times.

Search engines are nowadays widely applied to store and analyze logs generated by large-scale distributed systems. To adapt to various workload scenarios, log search engines such as Elasticsearch usually expose a large number of performance-related configuration parameters. As manual configuring is time consuming and labor intensive, automatically tuning configuration parameters to optimize performance has been an urgent need. However, it is challenging because: 1) Due to the complex implementation, the relationship between performance and configuration parameters is difficult to model and thus the objective function is actually a black box; 2) In addition to application parameters, JVM and kernel parameters are also closely related to the performance and together they construct a high dimensional configuration space; 3) To iteratively search for the best configuration, a tool is necessary to automatically deploy the newly generated configuration and launch tests to measure the corresponding performance. To address these challenges, this paper designs and implements HDConfigor, an automatic holistic configuration parameter tuning tool for log search engines. In order to solve the high dimensional optimization problem, we propose a modified Random EMbedding Bayesian Optimization algorithm (mREMBO) in HDConfigor which is a black-box approach. Instead of directly using a black-box optimization algorithm such as Bayesian optimization (BO), mREMBO first generates a lower dimensional embedded space through introducing a random embedding matrix and then performs BO in this embedded space. Therefore, HDConfigor is able to find a competitive configuration automatically and quickly. We evaluate HDConfigor in an Elasticsearch cluster with different workload scenarios. Experimental results show that compared with the default configuration, the best relative median indexing results achieved by mREMBO can reach $2.07\times $ . In addition, under the same number of trials, mREMBO is able to find a configuration with at least a further 10.31% improvement in throughput compared to Random search, Simulated Annealing and BO.

Nowadays, the widely used NoSQL databases play a fundamental role for big data storage, processing and analysis. However, NoSQL databases usually consume a large amount of computing resources and always run in a low performance state, which seriously impacts the end-to-end quality of service. Therefore, tuning the performance of NoSQL databases via benchmarking tools becomes a critical issue. Unfortunately, it is non-trivial to solve this problem because there exists a considerable number of performance-related configuration parameters from both the databases (application-specific) and the operating system kernel (kernel-specific), and manually tuning methods only consider a small subset of the whole candidate configuration parameters space, leading to a sub-optimal performance. To address this challenge, we design ConfAdvisor, an automatic configuration tuning framework for NoSQL database benchmarking. Specifically, ConfAdvisor treats the database performance as a black-box function of configuration parameters and leverage an online learning method to search the best configurations. Experimental results based on several popular NoSQL databases show that: 1) Through only tuning the kernel-specific parameters can improve database performance by up to 30%, which is often overlooked in previous studies. 2) ConfAdvisor is able to achieve well-tuned configurations with a very few trials, even in a high-dimension configuration space. And compared with default configurations, the well-tuned parameters can improve database performance by up to 88%.

In order to simultaneously power and cool hundreds of thousands of servers, large-scale data centers usually consume several to tens of megawatts of electricity. This enormous electricity consumption leads to considerable concerns in the electricity cost including both electricity bills and carbon tax. To achieve a sustainable data center, many Internet service providers begin to build their own on-site renewable energy plants to help reduce the electricity cost. However, considering the performance constraint of delay tolerant workloads and the lack of future information about the time-varying electricity price, carbon emission rate, and available on-site renewable energy, it is a fairly challenging problem that how to schedule the delay tolerant workloads to reduce the electricity cost of a sustainable data center. To address this challenging optimization problem, this paper proposes an online workload scheduling algorithm CECM based on the Lyapunov optimization framework, which is able to tradeoff between the electricity cost and the performance of delay tolerant workloads without any future information about the time-varying system states. With extensive simulations based on the real-life traces, we show that CECM is able to reduce the electricity cost by 9.26 percent, while still guaranteeing the performance constraint of delay tolerant workloads.