Wenqian Cui, Dianzhi Yu, Xiaoqi Jiao, Ziqiao Meng, Guangyan Zhang, Qichao Wang, Steven Y. Guo, Irwin King. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Zoned Namespace (ZNS) SSDs have emerged as a promising solution for eliminating device-level garbage collection and achieving predictable performance through the zone abstraction, which supports append-only writes. However, integrating ZNS SSDs into RAID configurations presents challenges, particularly concerning partial parity updates (PPUs) due to the no-overwrite property of ZNS SSDs. Existing solutions introduce dedicated metadata zones to manage PPUs, but these zones become performance bottlenecks, especially under high PPU workloads.In this paper, we introduce Zebra, a novel architecture for ZNS RAID that leverages the Zone Random Write Area (ZRWA) feature of modern ZNS SSDs. Our key insight is that the SSD’s write buffer (or ZRWA) in front of each zone is ideal for PPUs that exhibit a repeated sequential overwrite I/O pattern. By ensuring that the parity chunk fits within the ZRWA window, Zebra enables efficient PPUs within the ZRWA. Combined with a set of techniques, Zebra ensures both high performance and reliability. We evaluate Zebra on both large-zone and small-zone ZNS SSDs, two typical ZNS models. Our experiments with micro and application benchmarks show that Zebra improves the throughput by 1.1×-4.2× compared to RAIZN, a state-of-the-art ZNS RAID system.

Tokenising continuous speech into sequences of discrete tokens and modelling them with language models (LMs) has led to significant success in text-to-speech (TTS) synthesis. Despite these models can generate speech with high quality and naturalness, their synthesised samples can still suffer from artefacts, mispronunciation, word repeating, etc. In this paper, we argue these undesirable properties could partly be caused by the randomness of sampling-based strategies during the autoregressive decoding of LMs. Therefore, we look at maximization-based decoding approaches and propose Temporal Repetition Aware Diverse Beam Search (TRAD-BS) to find the most probable sequences of the generated speech tokens. Experiments with two recent LM-based TTS models demonstrate that our proposed maximisation-based decoding strategy generates speech with fewer mispronunciations and improved speaker consistency1.

This paper introduces PFlow-VC, a conditional flow matching voice conversion model that leverages fine-grained discrete pitch tokens and target speaker prompt information for expressive voice conversion (VC). Previous VC works primarily focus on speaker conversion, with further exploration needed in enhancing expressiveness (such as prosody and emotion) for timbre conversion. Unlike previous methods, we adopt a simple and efficient approach to enhance the style expressiveness of voice conversion models. Specifically, we pretrain a self-supervised pitch VQVAE model to discretize speaker-irrelevant pitch information and leverage a masked pitch-conditioned flow matching model for Mel-spectrogram synthesis, which provides in-context pitch modeling capabilities for the speaker conversion model, effectively improving the voice style transfer capacity. Additionally, we improve timbre similarity by combining global timbre embeddings with time-varying timbre tokens. Experiments on unseen LibriTTS test-clean and emotional speech dataset ESD show the superiority of the PFlow-VC model in both timbre conversion and style transfer. Audio samples are available on the demo page https://speechai-demo.github.io/PFlow-VC/.

Large organizations emit terabytes of logs every day in their cloud environment. Efficient data science on these logs via text search is crucial for gleaning operational insights and debugging production outages. Current log management systems either perform full-text indexing on a cluster of dedicated servers to provide efficient search at the expense of high storage cost, or store unindexed compressed logs on object storage at the expense of high search cost. We propose LogCloud, a new object-storage based log management system that supports both cheap compressed log storage and efficient search. LogCloud constructs inverted indices on compressed logs using a novel FM-index implementation that supports efficient querying from object storage directly, removing the need for dedicated indexing servers. Experiments on five public and five production log datasets show that LogCloud can achieve both cheap storage and search, scaling to TB-scale datasets.

Flash-based solid-state drives (SSDs) have been widely adopted in various storage systems, manifesting better performance than their forerunner HDDs. However, the characteristics of flash media post some drawbacks when deploying SSD-based storage systems. First, flash media have limited program/erase cycles, making them vulnerable to media failures. Second, SSD foreground I/Os can suffer from inconsistent performance due to interference from background operations like garbage collection (GC). The major solution to the above problems is to introduce data redundancy. Redundant data can not only detect raw bit errors and recover lost data but also enable I/O scheduling to sidestep SSDs that are under performance degradation. Compared with multi-replica, data coding is a more space-efficient way to provide redundancy. However, it is more challenging to simultaneously achieve low access latency, consistent performance, and fast recovery. This article examines the design of coded storage in existing storage systems, with a focus on flash storage systems, and how they address these challenges. The coded storage techniques are categorized into in-device coding, cross-device coding, and cross-machine coding. They are designed for different scenarios and purposes but share some design rationales in common. For each type of coded storage, we begin by presenting the theoretical bases, followed by an overview of how existing studies address the performance and endurance issues of coded storage from a systemic perspective. Finally, we review the history of coded storage, list several key insights from existing works, and speculate some promising directions for flash-oriented coded storage systems.

Silent Data Corruption (SDC) in processors can lead to various application-level issues, such as incorrect calculations and even data loss. Since traditional techniques are not effective in detecting these errors, it is very hard to address problems caused by SDCs in processors. For the same reason, knowledge about these SDCs in the wild is limited. In this article, we conduct an extensive study on CPU SDCs in a large production CPU population, encompassing over one million processors. In addition to collecting overall statistics, we perform a detailed study to understand (1) whether certain processor features are particularly vulnerable and their potential impacts on applications; (2) the reproducibility of CPU SDCs and the triggering conditions (e.g., temperature) of those less reproducible SDCs; and (3) the challenges to mitigate and handle CPU SDCs. We further investigate the implications that our observations obtained from the above researches have on the SDC fault models, SDC mitigation strategies, and the future research fields. In addition, we design an efficient SDC mitigation approach called Farron, which uses prioritized testing to detect highly reproducible SDCs and temperature control to mitigate less-reproducible SDCs. Our experimental results indicate that Farron can achieve better coverage of CPU SDCs with lower overall overhead, compared to the baseline used in Alibaba Cloud. This demonstrates that our observations are able to assist in SDC mitigation.

Cloud logs can be categorized into on-line, off-line, and near-line logs based on the access frequency. Among them, near-line logs are mainly used for debugging, which means they prefer a low query latency for better user experience. Besides, the storage system for near-line logs prefers a low overall cost including the storage cost to store compressed logs, and the computation cost to compress logs and execute queries. These requirements pose challenges to achieving fast and cheap cloud log storage. This article proposes LogGrep, the first log compression and query tool that exploits both static and runtime patterns to properly structurize and organize log data in fine-grained units. The key idea of LogGrep is “vertical partitioning”: it stores each log entry into multiple partitions by first parsing logs into variable vectors according to static patterns and then extracting runtime pattern(s) automatically within each variable vector. Based on such runtime patterns, LogGrep further decomposes the variable vectors into fine-grained units called “Capsules” and stamps each Capsule with a summary of its values. During the query process, LogGrep can avoid decompressing and scanning Capsules that cannot match the keywords, with the help of the extracted runtime patterns and the Capsule stamps. We further show that the interactive debugging can well utilize the advantages of the vertical-partitioning-based method and mitigate its weaknesses as well. To this end, LogGrep integrates incremental locating and partial reconstruction to mitigate the read amplification incurred by vertical-partitioning-based method. We evaluate LogGrep on 37 cloud logs from the production environment of Alibaba Cloud and the public datasets. The results show that LogGrep can reduce the query latency and the overall cost by an order of magnitude compared with state-of-the-art works. Such results have confirmed that it is worthwhile applying a more sophisticated vertical-partitioning-based method to accelerate queries on compressed cloud logs.

Hyper-parameter tuning (HPT) for deep learning (DL) models is prohibitively expensive. Sequential model-based optimization (SMBO) emerges as the state-of-the-art (SOTA) approach to automatically optimize HPT performance due to its heuristic advantages. Unfortunately, focusing on algorithm optimization rather than a large-scale parallel HPT system, existing SMBO-based approaches still cannot effectively remove their strong sequential nature, posing two performance problems: (1)extremely low tuning speedand (2)sub-optimal model quality. In this paper, we propose FastTuning, a fast, scalable, and generic system aiming at parallelly accelerating SMBO-based HPT for large DL/ML models. The key is to partition the highly complex search space into multiple smaller sub-spaces, each of which is assigned to and optimized by a different tuning worker in parallel. However, determining the right level of resource allocation to strike a balance between quality and cost remains a challenge. To address this, we further propose NIMBLE, a dynamic scheduling strategy that is specially designed for FastTuning, including (1) Dynamic Elimination Algorithm, (2) Sub-space Re-division, and (3) Posterior Information Sharing. Finally, we incorporate 6 SOTAs (i.e., 3 tuning algorithms and 3 parallel tuning tools) into FastTuning. Experimental results, on ResNet18, VGG19, ResNet50, and ResNet152, show that FastTuning can consistently offer much faster tuning speed (up to$80\times$) with better accuracy (up to 4.7% improvement), thereby enabling the application of automatic HPT to real-life DL models.

In cloud systems, near-line logs are mainly used for debugging, which means they prefer a low query latency for a better user experience, and like any other logs, they also prefer a low overall cost including storage cost to store compressed logs and computation cost to compress logs and execute queries.

Silent Data Corruption (SDC) in processors can lead to various application-level issues, such as incorrect calculations and even data loss. Since traditional techniques are not effective in detecting processor SDCs, it is very hard to address problems caused by SDCs. For the same reason, knowledge about SDCs in the wild is limited.

Compared with traditional solid-state drives (SSDs), open-channel SSDs (OCSSDs) expose their internal physical layout and provide a host-based flash translation layer (FTL) that allows host-side software to control the internal operations such as garbage collection (GC) and input/output (I/O) scheduling. In this paper, we comprehensively survey research works built on OCSSDs in recent years. We show how they leverage the features of OCSSDs to achieve high throughput, low latency, long lifetime, strong performance isolation, and high resource utilization. We categorize these efforts into five groups based on their optimization methods: adaptive interface customizing, rich FTL co-designing, internal parallelism exploiting, rational I/O scheduling, and efficient GC processing. We discuss the strengths and weaknesses of these efforts and find that almost all these efforts face a dilemma between performance effectiveness and management complexity. We hope that this survey can provide fundamental knowledge to researchers who want to enter this field and further inspire new ideas for the development of OCSSDs.

存储系统是计算机的核心,在人工智能、大数据、云计算和物联网等新兴战略产业的可持续发展中起着重要作用。随着处理器和网络设备性能不断提高,存储软件栈成为限制数据密集型系统性能的主要因素。近年来,新型存储设备因其打破“内存墙”的能力而受到广泛关注。这些设备包括支持块寻址的闪存设备、支持字节寻址的非易失性存储器、存算一体化设备以及大容量光存储。构建高吞量、低延迟和高可靠性的大规模存储系统,需要对算法、软件设计和硬件的持续创新。这些创新可以应对大规模、高性能复杂结构系统构建中存在的挑战,还可以增加相关系统的构建和应用经验,加快大数据处理系统的开发速度。 研究人员一直致力于解决“内存墙”问题,并改进相关软硬件生态系统,从而在新型存储系统设计和应用方面取得很大进展,包括但不限于以下方面:

The capability of generating speech with a specific type of emotion is desired for many human-computer interaction applications. Cross-speaker emotion transfer is a common approach to generating emotional speech when speech data with emotion labels from target speakers is not available for model training. This paper presents a novel cross-speaker emotion transfer system named iEmoTTS. The system is composed of an emotion encoder, a prosody predictor, and a timbre encoder. The emotion encoder extracts the identity of emotion type and the respective emotion intensity from the mel-spectrogram of input speech. The emotion intensity is measured by the posterior probability that the input utterance carries that emotion. The prosody predictor is used to provide prosodic features for emotion transfer. The timbre encoder provides timbre-related information for the system. Unlike many other studies which focus on disentangling speaker and style factors of speech, the iEmoTTS is designed to achieve cross-speaker emotion transfer via disentanglement between prosody and timbre. Prosody is considered the primary carrier of emotion-related speech characteristics, and timbre accounts for the essential characteristics for speaker identification. Zero-shot emotion transfer, meaning that the speech of target speakers is not seen in model training, is also realized with iEmoTTS. Extensive experiments of subjective evaluation have been carried out. The results demonstrate the effectiveness of iEmoTTS compared with other recently proposed systems of cross-speaker emotion transfer. It is shown that iEmoTTS can produce speech with designated emotion types and controllable emotion intensity. With appropriate information bottleneck capacity, iEmoTTS is able to transfer emotional information to a new speaker effectively. Audio samples are publicly available.

In the development of neural text-to-speech systems, model pre-training with a large amount of non-target speakers’ data is a common approach. However, in terms of ultimately achieved system performance for target speaker(s), the actual benefits of model pre-training are uncertain and unstable, depending very much on the quantity and text content of training data. This study aims to understand better why and how model pre-training can positively contribute to TTS system performance. It is postulated that the pre-training process plays a critical role in learning text-related variation in speech, while further training with the target speaker’s data aims to capture the speaker-related variation. Different test sets are created with varying degrees of similarity to target speaker data in terms of text content. Experiments show that leveraging a speaker-independent TTS trained on speech data with diverse text content can improve the target speaker TTS on domain-mismatched text. We also attempt to reduce the amount of pre-training data for a new text domain and improve the data and computational efficiency. It is found that the TTS system could achieve comparable performance when the pre-training data is reduced to 1/8 of its original size.

This study aims at designing an environment-aware text-tospeech (TTS) system that can generate speech to suit specific acoustic environments.It is also motivated by the desire to leverage massive data of speech audio from heterogeneous sources in TTS system development.The key idea is to model the acoustic environment in speech audio as a factor of data variability and incorporate it as a condition in the process of neural network based speech synthesis.Two embedding extractors are trained with two purposely constructed datasets for characterization and disentanglement of speaker and environment factors in speech.A neural network model is trained to generate speech from extracted speaker and environment embeddings.Objective and subjective evaluation results demonstrate that the proposed TTS system is able to effectively disentangle speaker and environment factors and synthesize speech audio that carries designated speaker characteristics and environment attribute.Audio samples are available online for demonstration 1 .

In-memory systems with erasure coding (EC) enabled are widely used to achieve high performance and data availability. However, as the scale of clusters grows, the server-level fail-slow problem is becoming increasingly frequent, which can create long tail latency. The influence of long tail latency is further amplified in EC-based systems due to the synchronous nature of multiple EC sub-operations. In this paper, we propose an EC-enabled in-memory storage system called Short Tail, which can achieve consistent performance and low latency for both reads and writes. First, Short Tail uses a lightweight request monitor to track the performance of each memory node and identify any fail-slow node. Second, Short Tail selectively performs degraded reads and redirected writes to avoid accessing fail-slow nodes. Finally, Short Tail posts an adaptive write strategy to reduce write amplification of small writes. We implement Short Tail on top of Memcached and compare it with two baseline systems. The experimental results show that Short Tail can reduce the P99 tail latency by up to 63.77%; it also brings significant improvements in the median latency and average latency.

This paper describes a novel design of a neural network-based speech generation model for learning prosodic representation.The problem of representation learning is formulated according to the information bottleneck (IB) principle.A modified VQ-VAE quantized layer is incorporated in the speech generation model to control the IB capacity and adjust the balance between reconstruction power and disentangle capability of the learned representation.The proposed model is able to learn word-level prosodic representations from speech data.With an optimized IB capacity, the learned representations not only are adequate to reconstruct the original speech but also can be used to transfer the prosody onto different textual content.Extensive results of the objective and subjective evaluation are presented to demonstrate the effect of IB capacity control, the effectiveness, and potential usage of the learned prosodic representation in controllable neural speech generation.

It is desirable to deploy a ConvNet system with high inference accuracy, as well as fast training and small inference memory. However, existing approaches to hyper-parameter tuning only focus on high accuracy. Although achieving high accuracy, tuning poorly can significantly increase the performance burden, and thus degrade the overall performance of a ConvNet system. In this article, we propose SmartTuning, an approach to identifying the hyper-parameters of a ConvNet system for high training speed and small working memory, with the restriction of high inference accuracy. The key idea of SmartTuning is to build a new performance model for a ConvNet system, and to integrate Bayesian Optimization to learn the relationship between the overall performance and the hyper-parameters of a ConvNet system. In this way, SmartTuning can balance inference accuracy, training speed and inference memory usage during the tuning process, and thus maximizes the overall performance of a ConvNet system. Our experiments show that SmartTuning can stably identify the hyper-parameter sets that offer very close accuracy with faster training speed (i.e., 7×-11× over MNIST and 2×-3× over CIFAR-10) and much less inference memory usage (i.e., 17×-23× over MNIST and 4×-9× over CIFAR-10), compared with existing tuning approaches.

As data progressively grows within data centers, the cloud storage systems continuously facechallenges in saving storage capacity and providing capabilities necessary to move big data within an acceptable time frame. In this paper, we present the Boafft, a cloud storage system with distributed deduplication. The Boafft achieves scalable throughput and capacity usingmultiple data servers to deduplicate data in parallel, with a minimal loss of deduplication ratio. Firstly, the Boafft uses an efficient data routing algorithm based on data similarity that reduces the network overhead by quickly identifying the storage location. Secondly, the Boafft maintains an in-memory similarity indexing in each data server that helps avoid a large number of random disk reads and writes, which in turn accelerates local data deduplication. Thirdly, the Boafft constructs hot fingerprint cache in each data server based on access frequency, so as to improve the data deduplication ratio. Our comparative analysis with EMC's stateful routing algorithm reveals that the Boafft can provide a comparatively high deduplication ratio with a low network bandwidth overhead. Moreover, the Boafft makes better usage of the storage space, with higher read/write bandwidth and good load balance.

Conventional RAID solutions with fixed layouts partition large disk enclosures so that each RAID group uses its own disks exclusively. This achieves good performance isolation across underlying disk groups, at the cost of disk under-utilization and slow RAID reconstruction from disk failures. We propose RAID+, a new RAID construction mechanism that spreads both normal I/O and reconstruction workloads to a larger disk pool in a balanced manner. Unlike systems conducting randomized placement, RAID+ employs deterministic addressing enabled by the mathematical properties of mutually orthogonal Latin squares, based on which it constructs 3-D data templates mapping a logical data volume to uniformly distributed disk blocks across all disks. While the total read/write volume remains unchanged, with or without disk failures, many more disk drives participate in data service and disk reconstruction. Our evaluation with a 60-drive disk enclosure using both synthetic and real-world workloads shows that RAID+ significantly speeds up data recovery while delivering better normal I/O performance and higher multi-tenant system throughput.

Hidden Markov models (HMMs) have a long tradition in automatic speech recognition (ASR) due to their capability of capturing temporal dynamic characteristics of speech. For emotion recognition from speech, three HMM based architectures are investigated and compared throughout the current paper, namely, the Gaussian mixture model based HMMs (GMM-HMMs), the subspace based Gaussian mixture model based HMMs (SGMM-HMMs) and the hybrid deep neural network HMMs (DNN-HMMs). Extensive emotion recognition experiments are carried out on these three architectures on the CASIA corpus, the Emo-DB corpus and the IEMOCAP database, respectively, and results are compared with those of state-of-the-art approaches. These HMM based architectures prove capable of constituting an effective model for speech emotion recognition. Also, the modeling accuracy is further enhanced by incorporating various advanced techniques from the ASR area. In particular, among all of the architectures, the SGMM-HMMs achieve the best performance in most of the experiments.

Disk-based graph systems store part or all of graph data on external devices like hard drives or SSDs, achieving scalability without excessive hardware. However, massive expensive disk I/Os remain the major performance bottleneck of disk-based graph processing. In this paper, we propose Redio, a new approach to accelerating disk-based graph processing by reducing disk I/Os. First, Redio observes that it is feasible to accommodate all vertex states in main memory and this can eliminate almost all vertex-related disk I/Os. Second, Redio introduces a dynamic selective scheduling scheme to identify inactive edges in each iteration and skip them when and only when such skipping can bring performance benefit. To improve its effectiveness, Redioin corporates a compact edge storage to improve data locality and an indexed bitmap to minimize its memory and computation overheads. We have implemented a single-node prototype for Redio under the edge-centric computation model. Extensive experiments show that Redio consistently outperforms well-known edge-centric disk-based systems in all experiments, delivering an average speedup of$4.33\times$on HDDs and$5.33\times$on SSDs over the fastest among them (i.e., GridGraph). Experimental results also show that Redio delivers an average speedup of$3.13\times$on HDDs and$1.28\times$on SSDs over the fastest among representative vertex-centric disk-based systems (i.e., FlashGraph).

Existing GPU-based approaches cannot yet meet the performance requirement for training very large convolutional neural networks (CNNs), where convolutional layers (Conv-layers) dominate the training time. In this paper, we find that no single convolution policy can always perform the fastest across all the computing phases. Then, we propose an approach called HyConv to accelerating multi-phase CNN computation by fine-grained policy selection. HyConv encapsulates existing convolution policies into a set of modules, and selects the fastest policy (a.k.a., winner policy) via one-round runtime measurement for computing each phase. Furthermore, HyConv uses a winner database to record the current winner policies, avoiding duplicate measurement later for the same parameter configuration. Our experimental results indicate that over all the used real-world CNN networks, HyConv consistently outperforms existing approaches on either a single GPU or four GPUs, with speedups of up to 3.3× and up to 1.6× over cuDNN-MM respectively. Such improvement can be explained by our result that HyConv delivers obviously better performance for most of single Conv-layers. Furthermore, HyConv has the ability to work with any parameter configuration and thus keeps better usability.

As one of the most important deep learning models, convolutional neural networks (CNNs) have achieved great successes in a number of applications such as image classification, speech recognition and nature language understanding. Training CNNs on large data sets is computationally expensive, leading to a flurry of research and development of open-source parallel implementations on GPUs. However, few studies have been performed to evaluate the performance characteristics of those implementations. In this paper, we conduct a comprehensive comparison of these implementations over a wide range of parameter configurations, investigate potential performance bottlenecks and point out a number of opportunities for further optimization.

Big graph computing can be performed over a single node, using recent systems such as GraphChi and XStream. Breadth-first graph search (a.k.a., BFS) has a pattern of marking each vertex only once as “visited” and then not using them in further computations. Existing single-server graph computing systems fail to take advantage of such access pattern of BFS for performance optimization, hence suffering from a lot of extra I/O latencies due to accessing no longer useful data elements of a big graph as well as wasting plenty of computing resources for processing them. In this paper, we propose FastBFS, a new approach that accelerates breadth-first graph search on a single server by leverage of the preceding access pattern during the BFS iterations over a big graph. First, FastBFS uses an edgecentric graph processing model to obtain the high bandwidth of sequential disk accesses without expensive data preprocessing. Second, with a new asynchronous trimming mechanism, FastBFS can efficiently reduce the size of a big graph by eliminating useless edges in parallel with the computation. Third, FastBFS schedules I/O streams efficiently and can attain greater parallelism if an additional disk is available. We implement FastBFS by modifying the X-Stream system developed by EPFL. Our experimental results show that FastBFS can outperform X-stream and GraphChi in the computing speed by up to 2.1 and 3.9 times faster respectively. With an additional disk, FastBFS can even outperform them by up to 3.6 and 6.5 times faster respectively.

With dramatic growth of data and rapid enhancement of computing powers, data accesses become the bottleneck restricting overall performance of a computer system. Emerging phase-change memory (PCM) is byte-addressable like DRAM, persistent like hard disks and Flash SSD, and about four orders of magnitude faster than hard disks or Flash SSDs for typical file system I/Os. The maturity of PCM from research to production provides a new opportunity for improving the I/O performance of a system. However, PCM also has some weaknesses, for example, long write latency, limited write endurance, and high active energy. Existing processor cache systems, main memory systems, and online storage systems are unable to leverage the advantages of PCM, and/or to mitigate PCM’s drawbacks. The reason behind this incompetence is that they are designed and optimized for SRAM, DRAM memory, and hard drives, respectively, instead of PCM memory. There have been some efforts concentrating on rethinking computer architectures and software systems for PCM. This article presents a detailed survey and review of the areas of computer architecture and software systems that are oriented to PCM devices. First, we identify key technical challenges that need to be addressed before this memory technology can be leveraged, in the form of processor cache, main memory, and online storage, to build high-performance computer systems. Second, we examine various designs of computer architectures and software systems that are PCM aware. Finally, we obtain several helpful observations and propose a few suggestions on how to leverage PCM to optimize the performance of a computer system.

Abstract The security implications of social bots are evident in consideration of the fact that data sharing and propagation functionality are well integrated with social media sites. Existing social bots primarily use Really Simple Syndication and OSN (online social network) application program interface to communicate with OSN servers. Researchers have profiled their behaviors well and have proposed various mechanisms to defend against them. We predict that a web test automation rootkit (WTAR) is a prospective approach for designing malicious social bots. In this paper, we first present the principles of designing WTAR‐based social bots. Second, we implement three WTAR‐based bot prototypes on Facebook, Twitter, and Weibo. Third, we validate this new threat by analyzing behaviors of the prototypes in a lab environment and on the Internet, and analyzing reports from widely‐used antivirus software. Our analyses show that WTAR‐based social bots have the following features: (i) they do not connect to OSN directly, and therefore produce few network flows; (ii) they can log in to OSNs easily and perform a variety of social activities; (iii) they can mimic the behaviors of a human user on an OSN. Finally, we propose several possible mechanisms in order to defend against WTAR‐based social bots. Copyright © 2016 John Wiley & Sons, Ltd.

Users of cloud storage usually assign different redundancy configurations (i.e.,$(k,m,w)$) of erasure codes, depending on the desired balance between performance and fault tolerance. Our study finds that with very low probability, one coding scheme chosen by rules of thumb, for a given redundancy configuration, performs best. In this paper, we propose CaCo, an efficient Cauchy coding approach for data storage in the cloud. First, CaCo uses Cauchy matrix heuristics to produce a matrix set. Second, for each matrix in this set, CaCo uses XOR schedule heuristics to generate a series of schedules. Finally, CaCo selects the shortest one from all the produced schedules. In such a way, CaCo has the ability to identify an optimal coding scheme, within the capability of the current state of the art, for an arbitrary given redundancy configuration. By leverage of CaCo's nature of ease to parallelize, we boost significantly the performance of the selection process with abundant computational resources in the cloud. We implement CaCo in the Hadoop distributed file system and evaluate its performance by comparing with “Hadoop-EC” developed by Microsoft research. Our experimental results indicate that CaCo can obtain an optimal coding scheme within acceptable time. Furthermore, CaCo outperforms Hadoop-EC by 26.68-40.18 percent in the encoding time and by 38.4-52.83 percent in the decoding time simultaneously.

The fast growing of data scale encourages the wide employment of data disks with large storage capacity. However, a mass of data disks' equipment will in turn increase the probability of data loss or damage, because of the appearance of various kinds of disk failures. To ensure the intactness of the hosted data, modern storage systems usually adopt erasure codes, which can recover the lost data by pre-storing a small amount of redundant information. As the most common case among all the recovery mechanisms, the single disk failure recovery has been receiving intensive attentions for the past few years. However, most of existing works still take the stripe-level recovery as their only consideration, and a considerable performance improvement on single failure disk reconstruction in the stack-level (i.e., a group of rotated stripes) is missed. To seize this potential improvement, in this paper we systematically study the problem of single failure recovery in the stack-level. We first propose two recovery mechanism based on greedy algorithm to seek for the near-optimal solution (BP-Scheme and STP-Scheme) for any erasure array code in stack level, and further design a rotated recovery algorithm (RR-Algorithm) to eliminate the size of required memory. Through a rigorous statistic analysis and intensive evaluation on a real system, the results show that BP-Scheme gains 3.4 to 38.9 percent (the average is 21.2 percent) higher recovery speed than Khan's Scheme and 3.4 to 34.8 percent (the average is 19.1 percent) higher recovery speed than Luo's U-Scheme, while STP-Scheme owns 3.4 to 46.9 percent (the average is 25.15 percent) and 3.4 to 41.1 percent (the average is 22.3 percent) higher recovery speed than Khan's Scheme and Luo's U-Scheme, respectively.

Disk additions to a RAID-6 storage system can simultaneously increase the I/O parallelism and expand the storage capacity. To regain a balanced load among both old and new disks, RAID-6 scaling requires moving certain data blocks onto newly added disks. Existing approaches to RAID-6 scaling are restricted by preserving a round-robin data distribution, and require migrating all the data, resulting in an expensive cost for RAID-6 scaling. In this paper, we propose Xscale, a new approach to accelerating X-code RAID-6 scaling by using lightweight data reorganization. Xscale minimizes the number of data blocks that require being moved, while maintaining a uniform data distribution across all disks. Furthermore, Xscale eliminates metadata updates while guaranteeing data consistency and data reliability. Compared with the round-robin approach, Xscale reduces the number of blocks to be moved by 63.689.5 percent, decreases the reorganization time by 35.62-37.26 percent, and reduces the I/O latency by 23.29-37.74 percent while the scaling programs are running in the background. In addition, there is no penalty in the performance of the data layout after scaling using Xscale, compared with the layouts maintained by other existing scaling approaches.

Disk additions to an RAID-6 storage system can increase the I/O parallelism and expand the storage capacity simultaneously. To regain load balance among all disks including old and new, RAID-6 scaling requires moving certain data blocks onto newly added disks. Existing approaches to RAID-6 scaling, restricted by preserving a round-robin data distribution, require migrating all the data, which results in an expensive cost for RAID-6 scaling. In this paper, we propose RS6—a new approach to accelerating RDP RAID-6 scaling by reducing disk I/Os and XOR operations. First, RS6 minimizes the number of data blocks to be moved while maintaining a uniform data distribution across all data disks. Second, RS6 piggybacks parity updates during data migration to reduce the cost of maintaining consistent parities. Third, RS6 selects parameters of data migration so as to reduce disk I/Os for parity updates. Our mathematical analysis indicates that RS6 provides uniform data distribution, minimal data migration, and fast data addressing. We also conducted extensive simulation experiments to quantitatively characterize the properties of RS6. The results show that, compared with existing “moving-everything” Round-Robin approaches, RS6 reduces the number of blocks to be moved by 60.0%–88.9%, and saves the migration time by 40.27%–69.88%.

Range-aggregate queries are to apply a certain aggregate function on all tuples within given query ranges. Existing approaches to range-aggregate queries are insufficient to quickly provide accurate results in big data environments. In this paper, we propose FastRAQ-a fast approach to range-aggregate queries in big data environments. FastRAQ first divides big data into independent partitions with a balanced partitioning algorithm, and then generates a local estimation sketch for each partition. When a range-aggregate query request arrives, FastRAQ obtains the result directly by summarizing local estimates from all partitions. FastRAQ has O(1) time complexity for data updates and O(N/P×B) time complexity for range-aggregate queries, where N is the number of distinct tuples for all dimensions, P is the partition number, and B is the bucket number in the histogram. We implement the FastRAQ approach on the Linux platform, and evaluate its performance with about 10 billions data records. Experimental results demonstrate that FastRAQ provides range-aggregate query results within a time period two orders of magnitude lower than that of Hive, while the relative error is less than 3 percent within the given confidence interval.

Disk additions to a RAID-4 storage system can increase the I/O parallelism and expand the storage capacity simultaneously. To regain load balance among all disks including old and new, RAID-4 scaling requires moving certain data blocks onto newly added disks. Existing data redistribution approaches to RAID-4 scaling, restricted by preserving a round-robin data distribution, require migrating all the data, which results in an expensive cost for RAID-4 scaling. In this paper, we propose McPod-a new data redistribution approach to accelerating RAID-4 scaling. McPod minimizes the number of data blocks to be moved while maintaining a uniform data distribution across all data disks. McPod also optimizes data migration with four techniques. First, it coalesces multiple accesses to physically successive blocks into a single I/O. Second, it piggybacks parity updates during data migration to reduce the cost of maintaining consistent parities. Third, it outsources all parity updates brought by RAID scaling to a surrogate disk. Fourth, it delays recording data migration on disks to minimize the number of metadata writes without compromising data reliability. We implement McPod in Linux Kernel 2.6.32.9, and evaluate its performance by replaying three real-system traces. The results demonstrate that McPod outperforms the existing “moving-everything” approach by 67.78-79.64 percent in redistribution time and by 14.24-27.16 percent in user response time. The experiments also illustrate that the performance of the RAID scaled using McPod is almost identical to that of the round-robin RAID.

In RAID-5, data and parity blocks are distributed across all disks in a round-robin fashion. Previous approaches to RAID-5 scaling preserve such round-robin distribution, therefore requiring all the data to be migrated. In this paper, we rethink RAID-5 data layout and propose a new approach to RAID-5 scaling called MiPiL. First, MiPiL minimizes data migration while maintaining a uniform data distribution, not only for regular data but also for parity data. It moves the minimum number of data blocks from old disks to new disks for regaining a uniform data distribution. Second, MiPiL optimizes online data migration with piggyback parity updates and lazy metadata updates. Piggyback parity updates during data migration reduce the numbers of additional XOR computations and disk I/Os. Lazy metadata updates minimize the number of metadata writes without compromising data reliability. We implement MiPiL in Linux Kernel 2.6.32.9, and evaluate its performance by replaying three real-system traces. The results demonstrate that MiPiL consistently outperforms the existing “moving-everything” approach by 74.07-77.57% in redistribution time and by 25.78-70.50% in user response time. The experiments also illustrate that under the WebSearch2 and Financial1 workloads, the performance of the RAID-5 scaled using MiPiL is almost identical to that of the round-robin RAID-5.

Scaling up a RAID-0 volume with added disks can increase its storage capacity and I/O bandwidth simultaneously. For preserving a round-robin data distribution, existing scaling approaches require all the data to be migrated. Such large data migration results in a long redistribution time as well as a negative impact on application performance. In this article, we present a new approach to RAID-0 scaling called FastScale. First, FastScale minimizes data migration, while maintaining a uniform data distribution. It moves only enough data blocks from old disks to fill an appropriate fraction of new disks. Second, FastScale optimizes data migration with access aggregation and lazy checkpoint. Access aggregation enables data migration to have a larger throughput due to a decrement of disk seeks. Lazy checkpoint minimizes the number of metadata writes without compromising data consistency. Using several real system disk traces, we evaluate the performance of FastScale through comparison with SLAS, one of the most efficient existing scaling approaches. The experiments show that FastScale can reduce redistribution time by up to 86.06% with smaller application I/O latencies. The experiments also illustrate that the performance of RAID-0 scaled using FastScale is almost identical to, or even better than, that of the round-robin RAID-0.

When a RAID-5 volume is scaled up with added disks, data have to be redistributed from original disks to all disks including the original and the new. Existing online scaling techniques suffer from long redistribution times as well as negative impacts on application performance. By leveraging our insight into a reordering window, this paper presents ALV, a new data redistribution approach to RAID-5 scaling. The reordering window is a result of the natural space hole as data being redistributed, and it grows in size. The data inside the reordering window can migrate in any order without overwriting other in-use data chunks. The ALV approach exploits three novel techniques. First, ALV changes the movement order of data chunks to access multiple successive chunks via a single I/O. Second, ALV updates mapping metadata lazily to minimize the number of metadata writes while ensuring data consistency. Third, ALV uses an on/off logical valve to adaptively adjust the redistribution rate depending on application workload. We implemented ALV in Linux Kernel 2.6.18 and evaluated its performance by replaying three real-system traces: TPC-C, Cello-99, and SPC-Web. The results demonstrated that ALV outperformed the conventional approach consistently by 53.31-73.91 percent in user response time and by 24.07-29.27 percent in redistribution time.

With the development of storage technology and applications, new caching policies are continuously being introduced. It becomes increasingly important for storage systems to be able to select the matched caching policy dynamically under varying workloads. This article proposes SOPA, a cache framework to adaptively select the matched policy and perform policy switches in storage systems. SOPA encapsulates the functions of a caching policy into a module, and enables online policy switching by policy reconstruction. SOPA then selects the policy matched with the workload dynamically by collecting and analyzing access traces. To reduce the decision-making cost, SOPA proposes an asynchronous decision making process. The simulation experiments show that no single caching policy performed well under all of the different workloads. With SOPA, a storage system could select the appropriate policy for different workloads. The real-system evaluation results show that SOPA reduced the average response time by up to 20.3% and 11.9% compared with LRU and ARC, respectively.

Out-of-band virtualization intrinsically has the potential to provide high performance and good scalability. Unfortunately, existing out-of-band virtualization systems have some limitations, such as restrictions to specific platforms and/or hardware. In this paper, we present a new out-of-band virtualization system, MagicStore, which is not limited to any specific hardware and supports three widely used host platforms: Windows, Solaris, and Linux. First, MagicStore uses the SLAS2 approach to scale round-robin striped volumes efficiently. Second, it survives panics and power failures robustly through a combination of lazy synchronizations, ordered writes, and REDO logging. Third, it also incorporates typical legacy storage quickly by analyzing partition tables and reconstructing logical volumes. Our evaluation results from representative experiments demonstrated that MagicStore has the ability to provide high performance, to introduce low processor overhead, and to have good scalability.

Round-robin striping, due to its uniform distribution and low-complexity computation, is widely used by applications which demand high bandwidth and massive storage. Because many systems are nonstoppable when their storage capacity and I/O bandwidth need increasing, an efficient and online mechanism to add more disks to striped volumes is very important. In this article, it is presented and proved that during data redistribution caused by scaling a round-robin striped volume, there is always a reordering window where data consistency can be maintained while changing the order of data movements. Furthermore, by exploiting the reordering window characteristic, SLAS is proposed to scale round-robin striped volumes, which reduces the cost of data redistribution effectively. First, SLAS applies a new mapping management solution based on a sliding window to support data redistribution without loss of scalability; second, it uses lazy updates of mapping metadata to decrease the number of metadata writes required by data redistribution; third, it changes the order of data chunk movements to aggregate reads/writes of data chunks. Our results from detailed simulations using real-system workloads show that, compared with the traditional approach, SLAS can reduce redistribution duration by up to 40.79% with similar maximum response time of foreground I/Os. Finally, our discussion indicates that the SLAS approach works for both disk addition and disk removal to/from striped volumes.