Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of Large Language Models (LLMs), and recent Mixture-of-Experts (MoE) extensions further enhance flexibility by dynamically combining multiple LoRA experts.However, existing MoE-augmented LoRA methods assume that experts operate independently, often leading to unstable routing, expert dominance.In this paper, we propose TalkLoRA, a communication-aware MoELoRA framework that relaxes this independence assumption by introducing expert-level communication prior to routing.TalkLoRA equips low-rank experts with a lightweight Talking Module that enables controlled information exchange across expert subspaces, producing a more robust global signal for routing.Theoretically, we show that expert communication smooths routing dynamics by mitigating perturbation amplification while strictly generalizing existing MoELoRA architectures.Empirically, TalkLoRA consistently outperforms vanilla LoRA and MoELoRA across diverse language understanding and generation tasks, achieving higher parameter efficiency and more balanced expert routing under comparable parameter budgets.These results highlight structured expert communication as a principled and effective enhancement for MoE-based parameterefficient adaptation.Code is available at https://github.com/why0129/TalkLoRA.

Learned index has emerged as a new indexing technique that leverages machine learning to accelerate in-memory data processing. However, current learned indexes are primarily designed to index numeric keys and lack robust support for variable-length string keys. In this paper, we propose a novel in-memory learned index called LISK (LearnedIndex forStringKeys) to support string keys. The novelty of LISK is two-fold. First, we propose a trie-like structure to address the limitations of linear models in fitting string keys. Each trie node indexes 8-byte key slices, which are organized as learned sub-indexes or B+-trees. Second, we present a new structure for learned sub-indexes, namely TLS (Two-layerLearnedSubindex), which is tailored to handle the complex distribution of string keys. TLS utilizes three key designs to improve the overall performance: (1) a two-phase hybrid index construction, (2) a second-derivative-based data partitioning, and (3) a cachefriendly overflow node design. We conduct extensive experiments on five datasets and six workloads to compare LISK with seven existing indexes, including five trie-based indexes and the state-of-the-art learned index LITS. The experimental results show that LISK achieves an average 1.99× (up to 7.87×) higher throughput across the six workloads on real-world datasets. Specifically, compared with LITS, LISK achieves an average 1.42× (up to 1.91× ) higher throughput. The source code of LISK is available athttps://github.com/suibianll/LISK/tree/master.

Improving software-managed cache efficiency is an important issue for various modern applications. Although LRU (Least Recently Used) has been widely used as the default replacement policy in many key-value caches, it suffers from low concurrency performance caused by frequent list updates. However, existing concurrency-aware schemes like FIFO sacrifice the LRU’s temporal locality of data access, leading to a drop in the cache’s hit ratio. In this work, aiming to improve the concurrency performance of cache management while maintaining a high hit ratio as LRU, we propose a novel concurrency-aware cache replacement policy called SEMU (SEgment-based MUlti-version cache replacement) that outperforms state-of-the-art schemes. The novelty of SEMU lies in two aspects. First, it uses an appendonly segment-based FIFO list to avoid locking the list, yielding high concurrency performance. Second, it adopts a multi-version mechanism to make the cache aware of the temporal locality of data access, which can offer a high hit ratio as LRU. With such designs, SEMU can offer high concurrency performance and a high hit ratio for the cache.We conduct extensive experiments on both synthetic and real world workloads to compare SEMU with a few existing schemes, including LRU, FIFO, 2Q, and FrozenHot. We also implement SEMU into a real key-value store, RocksDB, and perform a system-to-system evaluation. The performance comparison under the YCSB workloads shows that SEMU can efficiently improve the performance of RocksDB with the consideration of data concurrency. Particularly, in the multi-thread concurrent environment, SEMU averagely outperforms LRU, 2Q, FIFO, and FrozenHot by up to 1.9×, 1.6×, 80%, and 102% under the synthetic workloads and improves the throughput of RocksDB configured with default settings by up to 59.67% under the system-to-system comparison.

The Audio-Visual Question Answering (AVQA) task involves extracting question-related audio-visual clues from both temporal and spatial perspectives to answer questions accurately. Despite the promising performance of existing multi-modal AVQA models, thanks to large-scale pre-trained models, challenges remain in the field. Firstly, aligning audio-visual information across temporal and spatial dimensions is difficult. Secondly, the fusion of audio-visual information is often weighted inadequately, limiting model performance. To address the above issues, we design the Audio-Visual Adaptive Fusion Network (AVAF-Net), which uses contrastive learning to align audio-visual information temporally and spatially and adaptively adjusts fusion weights based on the question. Specifically, we initially align visual and audio information temporally through a temporal-alignment contrastive loss. This is followed by an audio-visual clue-mining module that highlights question-related cues, aligning them with the vocal region spatially using spatial alignment contrastive loss. Additionally, a question-oriented adaptive fusion module assigns different weights to audio and visual modalities based on the question content and then fuses them. The fused audio-visual cues are finally used to predict the answer. Extensive experiments on the MUSIC-AVQA dataset show that AVAF-Net surpasses all baseline models, with a maximum improvement of 15.90% in average accuracy and an average improvement of 9.80%.

Low-rank adaptation (LoRA) has been developed as an efficient approach for adapting large language models (LLMs) by finetuning two low-rank matrices, thereby reducing the number of trainable parameters.However, prior research indicates that many of the weights in these matrices are redundant, leading to inefficiencies in parameter utilization.To address this limitation, we introduce Dense Low-Rank Adaptation (DenseLoRA), a novel approach that enhances parameter efficiency while achieving superior performance compared to LoRA.DenseLoRA builds upon the concept of representation fine-tuning, incorporating a single Encoder-Decoder to refine and compress hidden representations across all adaptation layers before applying adaptation.Instead of relying on two redundant low-rank matrices as in LoRA, DenseLoRA adapts LLMs through a dense low-rank matrix, improving parameter utilization and adaptation efficiency.We evaluate DenseLoRA on various benchmarks, showing that it achieves 83.8% accuracy with only 0.01% of trainable parameters, compared to LoRA's 80.8% accuracy with 0.70% of trainable parameters on LLaMA3-8B.Additionally, we conduct extensive experiments to systematically assess the impact of DenseLoRA's components on overall model performance.Code is available at https://github.com/mulin-ahu/DenseLoRA.

Pre-trained medical vision foundation models have exhibited strong generalization across diverse tasks. However, the diversity of imaging modalities and segmentation requirements often leads to heterogeneous model architectures, making it difficult to adapt pre-trained weights across different model architectures. To address these limitations, we propose DU-Adapter, a parameter-efficient fine-tuning framework designed for cross-architecture adaptation of heterogeneous pre-trained weights. It features a densely connected U-shaped adapter with learnable convolutional downsampling and upsampling modules to align representations across heterogeneous models. Additionally, a depth-adjustable dense skip connection module allows flexible control over trainable parameter capacity to suit varying downstream segmentation tasks. Extensive experiments on polyp segmentation datasets show that DUAdapter outperforms both fully fine-tuned and lightweight tuning baselines, achieving superior performance with significantly reduced training cost.

Cache management is a critical concern for both key-value stores and relational DBMSs. The most significant challenge in cache management is the cache replacement strategy, which directly affects the throughput and latency of the cache manager. While the Least Recently Used (LRU) policy is widely adopted by many systems, it suffers from severe performance degradation in multi-threaded environments due to lock contention. This contention arises when multiple threads attempt to update the LRU list simultaneously. Motivated by this issue, we propose a new cache management scheme called twCache, designed to deliver high performance in concurrent environments. The novelty of twCache lies in two key aspects. First, it proposes to partition the replacement policy data structure into thread-wise sublists, each corresponding to one thread. Such a structure can enable thread isolation so that the requests from one thread will not introduce lock contention with other threads, yielding high concurrency performance. Second, we propose a low-cost technique to combine recency and hotness for victim selection during cache replacement. Each sublist is maintained as an LRU list, representing the recency of object requests. Each cached object's hot count is proposed to reflect its hotness, defined as the number of sublists visiting the object. We conducted extensive experiments to compare twCache with traditional algorithms (LRU, FIFO, and 2Q) and the state-of-the-art FrozenHot policy. Three types of trace are used, including 39 Twitter traces, 23 MSR traces, and 6 YCSB workloads. The results show that twCache achieves$12\times$and$7\times$higher throughputs than LRU on the Twitter and MSR traces, respectively. Meanwhile, twCache outperforms LRU by$4.8\times$in the average throughput under YCSB workloads.

Accurate disease grading (assessing lesion severity) and localization (pinpointing lesions) in CT images are critical for clinical diagnosis. While vision-language models typically require pixel-level annotations, medical CT annotation is timeconsuming, labor-intensive, and inconsistent due to differences in experience and annotation styles among experts. Consequently, most datasets provide only image-level category labels, posing significant challenges for current vision-language models under weak supervision, such as difficulty in accurately identifying ambiguous lesion features, susceptibility to confirmation bias, and the risk of error accumulation during training. To address these challenges, we propose CoWeT, a dual-branch collaborative weakly supervised text-guided visual-language model for disease grading and localization. Our model employs a dual-branch collaborative learning strategy, introducing mutual supervision between branches to refine features collaboratively and mitigate confirmation bias, alongside a novel Text Prompt Box (TPB) module that leverages pathology reports as textual prompts to intelligently filter pseudo-labels for robust lesion localization and grading under weak supervision. Evaluations on RTDP and LND datasets show that CoWeT outperforms state-of-the-art methods on accuracy and average precision.

Current methods for Music Emotion Recognition (MER) face challenges in effectively extracting features sensitive to emotions, especially those rich in temporal detail. Moreover, the narrow scope of music-related modalities impedes data integration from multiple sources, while including multiple modalities often leads to redundant information, which can degrade performance. To address these issues, we propose a lightweight framework for music emotion recognition that improves the extraction of features that are both sensitive to emotions and rich in temporal information and that integrates data from both audio and MIDI modalities while minimizing redundancy. Our approach involves developing two innovative unimodal encoders to learn embeddings from audio and MIDI-like features. Additionally, we introduce a Bi-modal Fusion Attention Model (BFAM) that integrates features from low-level to high-level semantic information across different modalities. Experimental evaluations on the EMOPIA and VGMIDI datasets show that our unimodal networks achieve accuracies that are 6.1% and 4.4% higher than baseline algorithms for MIDI and audio on the EMOPIA dataset, respectively. Furthermore, our BFAM achieves a 15.2% improvement in accuracy over the baseline, reaching 82.2%, which underscores its effectiveness for bi-modal MER applications.

Multimodal Summarization aims to use multimodal data to generate accurate and concise summaries for long sentences. While previous work has achieved promising success, they have overlooked the mismatching among multimodal semantics and lacked subject information guidance for adaptive referential images. Motivated by this observation, we propose ASSM, anAdaptiveSubject-focused modeling for multimodal summarization viaSemanticMatching. The novelty of ASSM lies in two aspects. First, we propose a multimodal semantic matching module that projects multimodal inputs into a shared joint embedding semantic space to determine whether the semantics between multimodalities are mismatching. Second, we propose an adaptive subject-focused guide module, which adaptively references images to learn subject tokens based on the multimodal semantic matching results. With these subject tokens, we are able to focus on the subject information, providing precise guidance for summary generation. We conduct extensive experiments on two standard benchmarks and compare ASSM with 17 existing models. The experimental results regarding ROUGE, BERTScore, and MoverScore show that the proposed ASSM model outperforms all competitors, achieving state-of-the-art performance and suggesting the effectiveness of our proposal. In addition, we provide a case study to further demonstrate the usability of ASSM.

In-memory learned index has been an efficient approach supporting in-memory fast data access. However, existing learned indexes are inefficient in supporting variable-length keys. To address this issue, we propose a new in-memory learned index called LIVAK that adopts a hybrid structure involving trie, learned index, and B+-tree. Each node indexes an 8-byte slice of keys, and we use learned indexes for large nodes but B+-trees for small nodes. Also, LIVAK presents a character re-encoding mechanism to avoid performance degradation. We compare LIVAK with B+-tree, Masstree, and SIndex on various datasets and workloads, and the results suggest the efficiency of LIVAK.

LSM-tree has been widely used in key-value stores to offer high write throughputs. However, LSM-tree suffers from the block-cache invalidation problem caused by periodical compaction operations, which lowers the efficiency of the block cache and leads to poor read performance, especially for range queries. To address this problem, we propose a novel cache component named Range Cache to accelerate range queries on LSM-based key-value stores. The differences between Range Cache and the traditional block cache lie in two aspects. First, Range Cache caches the query results, i.e., key-value pairs, rather than data blocks. Second, in contrast to the traditional block cache that utilizes a hash table to index data, Range Cache incorporates an ordered index, which is more efficient for range queries. Further, we integrate Range Cache into LSM-based key-value stores without disturbing other components. With Range Cache, we can eliminate the impact of compaction operations on the block cache, avoiding the block-cache invalidation problem and reducing disk I/Os for point/range queries. We implement Range Cache on top of RocksDB and conduct system-to-system comparisons to compare Range Cache with LevelDB, RocksDB, LSbM-tree, and RemixDB under various settings. The experimental results show that Range Cache can significantly improve the cache efficiency and increase the throughput, especially for range queries.

The feature alignment methods currently employed in domain adaptation for enhancing model transfer ability tend to compromise the discriminative capability of the model towards target features from different categories, consequently leading to a degradation in detector performance. To address this issue, this paper proposes a domain adaptation method for SAR (Synthetic Aperture Radar) image object detection based on instance contrastive learning. By utilizing a dual-branch prediction network, foreground and background instances are more accurately selected. Through instance contrastive learning, the proposed method narrows the distance between target features of the same category across different domains while widening the distance between target features of different categories. This approach facilitates alignment of the distributions of target features of the same category while preserving the discriminative nature of target features across different categories. Experimental results demonstrate that the proposed model achieves respective improvements in the Average Precision (AP) metric on two domain adaptation object detection datasets from optical images to SAR remote sensing images compared to SOTA methods.

Learned indexes incorporate machine learning models to predict the locations of keys in the dataset. However, achieving accurate prediction is difficult due to the learning models’ inability to fully capture data distribution. Existing learned indexes used data partition or pre-set error bounds to solve this problem. However, data partition requires constructing a higher index structure, and pre-set error bounds have to train many learning models. In the paper, we propose a new Hierarchical Learned Index with High-Precision query on end nodes (HLIHP), which aims to achieve higher query precision with lower training cost. First, we propose a Precision Correction Model to correct the prediction results, which associates the predicted results with the real positions of the keys. Then, a lightweight learned model is used to construct the superstructure of the index, which can find the end nodes quickly with low training cost and effectively support the query operation. We conduct experiments on four datasets, including covid, genome, osm, and planet, to evaluate the performance of our proposal. The results show that compared to the five hierarchical learned indexes, RMI, PGM-index, XIndex, FINEdex, and ALEX, HLIHP shows on average 2.42×, 1.78×, 5.68×, 6.11×, and 1.48× higher throughput in lookup performance.

Multi-tenancy is a crucial criterion for cloud service providers since it enables them to share resources among tenants, thus reducing costs. As all tenants share the same buffer space in the cloud server, a tenant’s overall performance will be affected by other tenants. Thus, developing an efficient buffering scheme to ensure tenant isolation becomes an urgent need. Aiming to solve this problem, this paper proposes a novel buffering scheme for managing the shared buffer in a multi-tenant cloud database. We first present an SLA (Service Level Agreement)-based model to quantify the buffering performance of each tenant. Then, we propose MTRP (Multi-Tenant Replacement Policy) for multi-tenant scenarios. The novelty of MTRP lies in three aspects: (1) it partitions the buffer into logical zones to implement tenant isolation. Each zone manages the pages one tenant uses, but the zone space can be dynamically adjusted by buffer replacements; (2) it proposes a two-step replacement algorithm, including global replacement and local replacement, which can improve space efficiency and reduce the SLA cost; (3) it adopts a reinforcement learning model to select the most appropriate zone for page replacement. The experimental results on various multi-tenant workloads show that MTRP achieves a higher hit ratio and lower SLA costs than LRU, LFU, LRU-2, and LeCaR.

ZNS SSD (Zoned Namespace SSD) is a new type of SSD based on the Zoned Namespace, which manages storage space with zones while providing host-level flash management.While previous work has evaluated the device-level performance of ZNS SSDs, it remains unclear how ZNS SSDs will perform on real SQL DBMSs.Since SQL DBMSs dominate the database market, it is worth performing an experimental analysis of ZNS SSDs when used for real SQL DBMSs.However, currently, existing SQL DBMSs cannot support ZNS SSDs directly.Thus, we turn to MyRocks, which is a branch of MySQL but replaces the InnoDB engine with RocksDB.Next, we install a ZNScompatible file system module named ZenFS, which provides the FTL functionalities for ZNS SSDs, and configure RocksDB to run on ZenFS.With such a configuration, we set up a real SQL environment to evaluate the performance of ZNS SSDs.The experimental evaluation consists of two parts.First, we verify the effectiveness of ZenFS by measuring the write amplification factor and space amplification factor of ZenFS when running on a real ZNS SSD.Second, we test the performance of MyRocks on a real ZNS SSD.To compare ZNS SSDs with traditional SSDs, we also include two types of old SSDs: an NVMe SSD using the PCIe interface and a SATA SSD using the SATA interface.The experimental results regarding tps, latency, and multi-thread scalability show that ZNS SSDs outperform NVMe SSDs regarding write performance, especially for deletions and updates.Meanwhile, all SSDs show similar read performance.Additionally, ZNS SSDs exhibited better performance stability than NVMe SSDs when the space used increased.As a result, ZNS SSDs are more suitable for write-intensive applications.

ZNS SSD (Zoned Namespace SSD) is a new SSD storage device that was proposed recently and has attracted extensive attention from academia and industry.ZNS SSDs divide the internal storage space into many zones.Each zone can be read randomly but can only be written sequentially.In addition, the host needs to control some ZNS SSD operations, such as garbage collection.However, what is the performance of ZNS SSD on real database systems?This question remains unanswered.To answer this question, this paper chooses the popular key-value store, RocksDB, as the targeted system and tests the performance of ZNS SSD on RocksDB experimentally to provide a reference for the rational use of ZNS SSD in database systems in the future.A comprehensive evaluation shows that ZNS SSDs perform better when many compaction operations are involved.Moreover, ZNS SSDs do not show the same performance degradation as traditional SSDs under high workloads.Therefore, ZNS SSDs are more suitable for application scenarios that involve frequent data updates.

Nonvolatile memory (NVM) suffers from more serious nonuniform memory access (NUMA) effects than DRAM because of the lower bandwidth and higher latency. While numerous works have aimed at optimizing NVM indexes, only a few of them tried to address the NUMA impact. Existing approaches mainly rely on local NVM write buffers or DRAM-based read buffers to mitigate the cost of remote NVM access, which introduces memory overhead and causes performance degradation for lookup and scan operations. In this article, we present NOBtree, a new NUMA-optimized persistent tree index. The novelty of NOBtree is two-fold. First, NOBtree presents per-NUMA replication and an efficient node-migration mechanism to reduce remote NVM access. Second, NOBtree proposes a NUMA-aware NVM allocator to improve the insert performance and scalability. We conducted experiments on six workloads to evaluate the performance of NOBtree. The results show that NOBtree can effectively reduce the number of remote NVM accesses. Moreover, NOBtree outperforms existing persistent indexes, including TLBtree, Fast&Fair, ROART, and PACtree, by up to$3.23\times $in throughput and$4.07\times $in latency.

Key-value stores (KVStores), such as LevelDB and Redis, have been widely used in real-world production environments. To guarantee data durability and availability, traditional KVStores suffer from high write latency, mainly caused by the long network and data-persisting time. To solve this problem, this article presents a novel data-persisting path for KVStores, allowing remote clients to persist data to the KVStore server with$\mu s$-level latency. The novelty of this study is threefold. First, we propose PMRDirect, which utilizes a persistent memory region (PMR) in the NVM express standard to construct a direct data-persisting path from the RDMA networking card (NIC) to the PMR region inside an SSD. Second, to showcase PMRDirect in KVStores, we developed a new accessing stack called PMRAccess, enabling remote clients to access existing KVStores and providing durability for each write request. Specifically, we present a low-latency RDMA-based messaging mode and a chunk-based PMR management in PMRAccess to reduce write latency and improve system throughput. Finally, we conducted extensive experiments to evaluate the performance of our proposals. We first compared PMRDirect with a few remote data-persisting paths to show its effectiveness. Then, we evaluated PMRAccess upon two KVStores, including LibCuckoo (an in-memory KVStore) and LevelDB (an in-storage KVStore). The results showed that PMRAccess outperformed the SSD-based accessing stack by up to$6.1\times $in write throughput and$36\times $in write tail latency, and it achieved$1.7\times $higher write throughput and$0.59\times $lower write tail latency over the PMEM-based accessing stack. Further, we conducted a system-to-system comparison between the PMRAccess-integrated LibCuckoo and Redis, and the results showed our proposal achieved up to$13\times $higher throughputs and$40\times $lower write latency than Redis.

Colonoscopy has emerged as a crucial examination for early colorectal cancer (CRC) diagnosis. Early detection of polyps can significantly enhance the survival rate of colorectal cancer. Most recent weakly supervised methods for detecting polyps are based on multiple instance learning (MIL), which employs labeled training data at the video-level (bag-level) to identify polyps at the frame-level (instance-level). However, existing methods often use the same features without considering the differences of video and snippet. Video classification usually focuses more on global features, while snippet classification relies more on leveraging multi-granularity detail information. This paper proposes decoupling the MIL network into the feature encoder and instance decoder. Furthermore, we introduce a novel Snippet-wise Cross Fusion Attention (SCA) that captures rich temporal context semantic features for instance classification. Additionally, our approach incorporates a parameter-efficient finetuning architecture called convolutional adapters, which aims to enhance the training process stability and improve the model's performance. Experimental results demonstrate consistent improvements over state-of-the-art methods on a newly introduced large-scale colonoscopy video dataset by a considerable 7.9% AUC and 1.16% AP. Our code and dataset will be made publicly available at: https://github.com/kanydao/Decoupling-MIL.

Few-shot learning (FSL) is a machine learning method aimed at learning problem-solving models from a small number of samples. Few-shot classification is a task that utilizes few-shot learning for image classification. However, the existing models have poor classification performance when dealing with datasets with large differences in target scale distribution. In view of the large difference in the scale distribution of the sample target in the few-shot learning dataset, this paper proposes a novel few-shot classification model based on multi-scale Vision Transformer using semantic prompt. This model uses the Transformer block which has more generalization ability and long-range dependencies, and constructs a multi-scale feature fusion model on the large-scale cross-domain and cross-scale dataset Meta-Dataset which consists of 10 datasets, including natural image datasets, handwritten characters, and graffiti datasets. By using an explicit multi-scale feature fusion method, the model can obtain category features with cross scale information, thereby improving the robustness of the model to sample targets of different scales. In addition, the model improved its performance by learning additional semantic prompt, achieving a 2.43% improvement.

In this paper, we propose a new space-efficient key-value store called ZoneKV for ZNS (Zoned Namespace) SSDs. We observe that existing work on adapting RocksDB to ZNS SSDs will cause fragmentation of zones and severe space amplification. Thus, we propose a lifetime-based zone storage model and a level-specific zone allocation algorithm to store SSTables with a similar lifetime in the same zone. We evaluate ZoneKV on a real ZNS SSD. The results show that ZoneKV can reduce up to 60% space amplification and maintain higher throughputs than two competitors, RocksDB and ZenFS.

So far, most LSM-tree-based storage engines adopt either leveling or tiering compaction. We note that while leveling compaction can deliver high search performance and low space amplification, it has a high rate of write amplification (therefore delivering poor write performance). On the other hand, tiering compaction has a low rate of write amplification (therefore delivering good write performance) but has poor search performance and high space amplification. Aiming to close the performance gap between leveling and tiering databases, this paper proposes a new storage engine called B+LSM. The novel ideas of B+LSM lie in two aspects: (1) B+LSM replaces the underlying level structure of LSM-tree with a B+-tree-like tree, and each tree node is defined as a Bundle Compaction Unit (BCU), whose size is allowed to be dynamically changed with workload statistics to balance read and write performance. (2) B+LSM proposes a new node-grained compaction scheme called Bundle Compaction. Bundle compaction is always triggered to merge all the data within a BCU node, partition them into bundles, and then send bundles to the children. Such a compaction scheme can take advantage of leveling and tiering compaction by auto-tuning the size of BCU nodes. We implemented B+LSM and compared it with LevelDB, RocksDB, PebblesDB, and L2SM on the YCSB workloads. The results show that B+LSM can achieve high time performance and reduce space amplification on both static and dynamic workloads.

Remote direct memory access (RDMA) supports zero-copy networking by transferring data from clients directly to host memory, eliminating the need to copy data between clients' memory and the data buffers in the hosting server. However, the hosting server must design efficient memory management schemes to handle incoming clients' data. In this paper, we propose a high-performance host memory management scheme called HM2 for RDMA-enabled distributed systems. We present a new buffer structure for incoming data from clients. In addition, we propose efficient data processing methods to reduce network transfers between clients and servers. We conducted a preliminary experiment to evaluate HM2, and the results showed HM2 achieved higher throughput than existing schemes, including L5 and FaRM.

The log-structured merge tree (LSM-tree) has been widely adopted as a backbone of modern key-value stores. However, the multiple exponentially increased levels of LSM-tree makes it suffer from high write amplification. Existing studies often improve the write performance by sacrificing the read performance, which is inefficient to make trade-offs between the update and search efficiency. In this paper, we exploit nonvolatile memory (NVM) to address the write amplification issue for systems with NVM-SSD hybrid storage, and further propose a reinforcement learning method to navigate between update and search efficiency on the varying workloads. Specifically, we first propose a lightweight hot data identification method to efficiently capture access recency as well as frequency in NVM with relative large capacity. On this basis, we can eliminate different versions of frequently updated data in high-performance NVM without pushing them to SSD. To improve the data access locality and facilitate fine-grained index tuning in each level, we devise a virtual-split method to partition the key space gradually without extra write amplification. Finally, we propose a cost based Q-learning algorithm to adaptively tune the data organizations of each partition according to the changing access patterns. Experimental results show that our approach outperforms existing methods by up to 2.67×.

Single-stage Human-Object Interaction (HOI) detection methods have attracted considerable attention due to their high efficiency. Existing methods tend to concentrate the detection of humans and objects in one decoder without considering the differences between them, which causes tremendous pressure on a single decoder and affects the detection effect. This paper aims to decouple the detection decoder of humans and objects. In particular, we advocate and propose a novel human-object decoupling network (HOD) that divides the decoder into three tasks: human detection, object detection, and action classification. The network uses the random erasure training strategy to improve the model’s generalization ability and introduces pose features to handle the long-tailed problem. In addition, we design a pose fusion branch to alleviate the semantic gap between pose and HOI datasets. The experimental results suggest that our method achieves consistent improvements over the state-of-theart across different datasets, utilizing only image information. Specifically for HICO-Det, our method outperforms existing methods by a large margin, with a significant relative mAP gain of 7.8%. Our source code will be publicly available upon acceptance.

Learned Indexes aim to use machine-learning models to predict the target addresses of requested data, which have been demonstrated efficient for in-memory data accesses. However, modern database systems mainly use flash-memory-based solid-state drives (SSDs) as storage devices, and current learned indexes fail to work on SSD-based databases. In this paper, we propose LIFM (Learned Index for Flash Memory), a new persistent and updatable learned index for flash-memory-based SSDs. Unlike existing learned indexes that aim to reduce memory access, LIFM is designed to reduce I/O costs. The novelty of LIFM lies in three aspects. First, LIFM proposes a hierarchical tree structure that combines the advantages of traditional B+-tree and in-memory learned indexes. Second, LIFM employs a model-based data placement scheme to reduce the page reads or writes. Third, LIFM uses a flash-memory-friendly storage layout to reduce additional access to flash memory. We conduct experiments on a real SSD and compare LIFM with B+-tree and two recently proposed learned indexes, ALEX and FITing-tree. The results suggest the efficiency of LIFM.

OLAP applications often perform aggregation queries, such as sum, avg, count, max, and min, to process a large dataset for getting a small result, and the key challenge for optimizing aggregation queries is to reduce the data volume needed to be read from disks. To address this problem, this paper presents an efficient caching component called FGCache (Fine-Grained Cache) for aggregation queries. FGCache is a query-driven caching policy, which aims to maintain the results of aggregation queries in the memory according to the predicates of user queries. The query results are divided into sets with fine-grained querying intervals, which are maintained via an in-memory tree index. We experimentally evaluate the performance of FGCache on various datasets and settings. The results suggest the efficiency and effectiveness of our proposal.

Zoned Namespace SSDs (ZNS SSDs) are a new kind of SSDs that can offer stabler writing performance and higher space efficiency than traditional SSDs. In this paper, we propose LeanKV, a novel LSM-tree-based key-value store optimized for ZNS SSDs, which aims to reduce the data migration cost of garbage collection. We note that existing ZNSSSD-aware data management approaches will lead to zone fragmentation in write-heavy workloads, causing extra data migration during garbage collection. To address this problem, we first present a new method to predict the lifetime of an SSTable based on its compaction history. Then, we propose a lifetime-based algorithm to cluster SSTables according to lifetime. The SSTables in the same cluster will be put into the same zone, ensuring the SSTables in each zone have similar lifetimes. With such a mechanism, less valid data is needed to be moved during garbage collection. We implement LeanKV on top of RocksDB and compare LeanKV with RocksDB and ZenFS on a real ZNS SSD (Western Digital ZN540). The results show that LeanKV can effectively reduce the amount of data migration and the number of files migrated during garbage collection under various workloads. Meanwhile, LeanKV achieves higher throughputs than RocksDB and ZenFS.

This paper proposes a novel learned index called DTtree that can optimize both read and write performance. DTtree adopts two novel designs. (1) To improve write performance, DTtree uses a dynamic filling rate and overflow buffers for each leaf node. The filling rate of a leaf node is dynamically determined according to the read/write tendency of the leaf node so as to improve space efficiency and reduce structure modification operations. In addition, DTtree uses an overflow buffer for a line in each leaf node. The overflow buffer is also dynamically allocated, and only write-intensive leaf nodes may have overflow buffers, reducing unnecessary space costs. (2) To improve read performance, we propose an in-node hot cache to cache hot records within a line of a leaf node, reducing additional memory access to the overflow buffer. As skewed access is common in real-world applications, using in-node hot caches can efficiently improve the read performance of DTtree. We experimentally evaluate DTtree on three datasets. The results on various workloads show that DTtree can reduce SMO operations effectively and achieve higher read/write performance than B+-tree and two representative learned indexes, PGM-Index and ALEX.

Object detectors trained with massive labeled data often suffer performance degradation in some particular scenarios with data distribution gap.Domain classifier is a commonly used method in the existing domain adaptation algorithms to alleviate the domain discrepancy.However, it has problems of instability in training, difficulty in obtaining optimal solutions and converging to an equilibrium point.To tackle this issue, we propose a novel balanced domain classifier (BDC), which not only eliminates domain discrepancy but also makes the domain classifier and the feature extractor maintain equilibrium during the adversarial learning.Furthermore, we propose an appropriate learning rate adjustment strategy, which makes the detection model converge to an equilibrium point more stably and more rapidly.Based on the domain-invariant region proposal network, we propose a cross-domain object detection model called Faster R-CNN with Balanced Domain Classifier (BDC-FR).The experimental results show that BDC-FR can effectively improve the performance of the cross-domain object detection model.

Hybrid memory composed of DRAM and persistent memory (PM) offers a promising way to realize large-capacity main memory supporting in-memory data storage and computing. However, traditional buffer management schemes focus on improving the hit ratio but lack awareness of the limitations of PM, e.g., slower write time and lower write endurance than DRAM. Therefore, developing new buffer management policies that can reduce costly write-backs of PM blocks while maintaining high performance for the hybrid buffer, is of paramount importance. Existing approaches mainly use a page-grained buffering policy, which will cause unnecessary data migrations between DRAM and PM, leading to a high number of disk I/Os and PM writes. Aiming to reduce I/O costs and PM writes, we propose a new buffer manager named HiBuffer for DRAM/PM-based hybrid memory systems. HiBuffer presents several novel ideas. First, it adopts multigrained data layouts to manage the hybrid buffer cooperatively. In addition to the page granularity, we introduce Lines for the DRAM buffer and Sectors to the PM buffer, forming a buffer with three granularities, including Line, Sector, and Page. We prove that the multigrained cooperative buffer management can deliver higher performance than existing page-grained schemes. Second, we propose a sector-grained method to migrate data from DRAM to PM, which can avoid unnecessary data movements and reduce PM writes. Third, we use an out-of-place updating mechanism to absorb updates in DRAM, which can further reduce the writes to PM. We compare HiBuffer with three existing schemes, including LRU, CLOCK-DWF, and MiniPage, on five synthetic workloads and the YCSB benchmark using real Intel Optane DC PM. The results in terms of various metrics, including running time, PM writes, hit ratio, and disk I/Os, suggest the efficiency of HiBuffer. In particular, HiBuffer reduces the running time by up to 37.8% and the writes to PM by up to 83% compared to the competitors when evaluated on the YCSB benchmark.

The emergence of Zoned Namespaces SSDs, known as ZNS SSDs, has introduced new challenges and opportunities for key-value stores. Since LSM-tree has been widely adopted in key-value stores, such as RocksDB and HBase. Compared with traditional SSDs, ZNS SSDs have the advantages of lower over-provisioning costs, higher and more stable performance, and better space efficiency. However, ZNS SSDs also have some limitations like sequential writes on zones. To make ZNS SSDs efficient in LSM-tree-based key-value stores, it is necessary to revisit the LSM-tree to make it adapt to ZNS SSDs. In this paper, we first analyze the challenges of adapting LSM-tree to ZNS SSDs. Then, we propose a high-level architecture for a ZNS-SSD-aware LSM-tree-based key-value store.

With the development of big data applications, online analytics systems that aim to offer decision support for various businesses and tasks have become a research focus, which calls for efficient approaches to handling OLAP queries. However, traditional OLAP systems suffer from the costly interactions with disks or SSDs, making them hard to deliver high performance for OLAP query processing. In this paper, we propose to use the emerging persistent memory to construct an efficient in-memory analytics system to improve the performance for OLAP query processing. We present the overall architecture as well as the detailed algorithms for the proposed system, which is named PM-Picker, and finally discuss the implementation issue of the system.

In recent years, weakly supervised object detection (WSOD) methods using only image-level labels have received increasing attention, Due to the difficulty of manually labeling large-scale remote sensing images. However, existing methods cannot generate high-quality proposals when applying proposal generation methods to RSIs. Meanwhile, these methods ignore the fact that there are a large number of objects of different scales in RSIs. To address these issues, we propose a unique end-to-end dynamic feature fusion network (DFFNet) for WSOD in RSIs. First, we propose an intersection-over-union selective search (IoU-SS) algorithm to generate high-quality proposals by preferentially merging regions with high IoU. Furthermore, we design a novel and flexible dynamic feature fusion (DFF) module to dynamically acquire features of objects at different scales based on the information of the input image. The performance of WSOD in RSIs is further improved by using high-quality proposals and dynamically fused features. Comprehensive experiments and comparisons with state-of-the-art methods on two datasets of RSIs, i.e., NWPU VHR-10.v2 and DIOR, demonstrate the superiority of our proposed method.

The emergence of new storage like persistent memory (PM) and zoned namespaces SSDs (ZNS-SSDs) introduces new challenges and opportunities for distributed key-value stores. Since LSM-tree has been widely adopted in distributed key-value stores, such as RocksDB and HBase, it is necessary to revisit the LSM-tree to make it adapt to new storage. In this paper, we first analyze the challenges of adapting the LSM-tree for new storage. Then, we propose a high-level architecture for a new-storage-aware LSM-tree-based key-value store called Hybrid-LSM. We explain the key structural issues of different storage layers in Hybrid-LSM and present some preliminary design ideas.

Cloud data storage relies on efficient cache systems to offer high performance for intensive reads/writes on big data. Due to the large data volume of cloud data storage and the limited capacity of DRAM, current cloud vendors prefer to use SSDs (Solid State Drives) but not DRAM to build the cache system. However, traditional SSDs have a serious over-provisioning problem and a high cost in garbage collection. Thus, the performance of SSD-based cache systems will drop quickly when the usage of SSDs increases. Recently, Zoned Namespaces (ZNS) SSDs have emerged as a hot topic in both academics and industries. Compared to conventional SSDs, ZNS SSDs have the advantages of less overhead of garbage collection and lower over-provisioning costs. Therefore, ZNS SSDs have been a better candidate for the cache system for cloud storage. However, ZNS SSDs only accept sequential writes, and the zones inside ZNS SSDs need to be carefully managed to maximize the advantages of ZNS SSDs. Therefore, making the cache system adapt to ZNS SSDs is becoming a challenging issue. In this paper, we demonstrate ZonedStore, a novel ZNS-aware cache system for cloud data storage. After a brief introduction to the architecture of ZonedStore, we present the key designs of ZonedStore, including a Zone Manager to control the space allocation and operations on ZNS SSDs, a Multi-Layer Buffer Manager, and an In-Memory Concurrent Index to accelerate accesses. Finally, we present a case study to demonstrate the working process and performance of ZonedStore.

Cloud services such as multi-tenant content delivery networks (CDN) have become a trend because they can offer personalized business support for users. However, as the business modes of different users are usually different, it is not appropriate to use a single cache strategy on cloud servers. First, a single cache cannot adapt to various access patterns of tenants. Second, a single cache will also affect the isolation among different users. In this paper, we propose MyCache, a new framework to deliver personalized cache management for multi-tenant cloud services. We first discuss the impact of access patterns, which motivates MyCache. Then, we briefly introduce the architecture of MyCache, and finally, we present preliminary experimental results to show the feasibility and superiority of our proposal.

LSM-tree has been widely used as a write-optimized storage engine in many key-value stores, such as LevelDB and RocksDB. However, conventional compaction operations on the LSM-tree need to read, merge, and write many SSTables, which we call Table Compaction in this paper. Table Compaction will cause two major problems, namely write amplification and block-cache invalidation. They will lower both write and read performance of the LSM-tree. To address these issues, we propose a novel compaction scheme named Block Compaction that adopts a block-grained merging policy to perform compaction operations on the LSM-tree. Block Compaction identifies the boundaries of data blocks and tries to avoid reusing data blocks, which not only reduces the write amplification but also alleviates the block-cache invalidation. We present cost analysis to theoretically demonstrate that Block Compaction is more efficient than the existing Table Compaction. Furthermore, we analyze the side-effects of Block Compaction and present three optimizations: (1) Selective Compaction is to reduce the space amplification of Block Compaction by integrating Table Compaction with Block Compaction. (2) Parallel Merging divides a compaction task into several sub-tasks and uses multiple workers to accomplish sub-tasks in parallel. (3) Lazy Deletion mitigates the overhead caused by traversing files at the tail of compaction operations. We implement a new key-value store named BlockDB based on Block Compaction and its optimizations. Then, we compare BlockDB with LevelDB, RocksDB, and L2SM using the YCSB benchmark. The results show that BlockDB can reduce write amplification up to 32% and running time by up to 43.6%, compared to its competitors. In addition, it can maintain the high performance for point lookups and range scans.

With the development of computer vision and the growth of image data volume, object detection is more and more widely used in people's life. Recently, object detection network base on transformer has good effect. However, it doesn't pay attention to objects' distinguishable features and the correlation of different distinguishable features, which makes its accuracy not satisfactory. In this paper, we propose a new transform correlation network (FTCN) based on rotational attention for object detection, which can be inserted into various modules of the object detection network theoretically. It can help the network to pay more attention to distinguishable features and their correlation, and increase the accuracy of the network. Compared to transformer, FTCN reduces the large computational cost and parameters that encoders and decoders require. Experimental results show that FTCN can increase the accuracy of the network to a certain extent in Microsoft COCO and Pascal VOC. The comparative results of GFLOPs and Param show that our approach has less computational cost and parameters compared with transformer methods.

Data diversity is one of the main characteristics of big data, which makes a growing number of multi-modal data. For example, in e-commerce applications, a product is often described simultaneously through text, image, video, etc. However, traditional text search engines are mainly oriented to text data and cannot offer high-performance indexing and search on multi-modal data. In this paper, we propose a high-performance hybrid index structure named MMH-index to enhance Apache Lucene, an open-source Java library providing powerful indexing and search features, with multi-modal indexing and searching. The MMH-index optimizes the classical inverted index in Lucene with an innovative design called modality bitmap. By adding modality bitmaps to the dictionary of the inverted index, MMH-index can reduce the redundant data in the index and the average length of the posting lists, thus decreasing the space consumption and the time cost of multi-modal queries. We implement MMH-index in Lucene and evaluate its performance on two real multi-modal datasets. The experimental results show that compared with the traditional inverted index in Lucene, MMH-index achieves significant improvement in space consumption and multi-modal query performance.

Buffer management is an essential technology for database management systems.Traditional buffer management employs an empirical approach based on access recency or frequency which fails to adapt to access-pattern changes in various database applications.In this paper, we present a new access-pattern-aware buffer manager called PBM (Personalized Buffer Manager), which can detect the access patterns for each database file and use a specific buffering policy for each database file.In particular, we propose a workload classifier to detect the access pattern of a database file.Then, we partition the buffer into various zones, set different sizes for each zone, and select the most suitable buffering scheme for each zone.With such a mechanism, each zone is responsible for caching a specific database file, and we can realize a personalized buffer manager for different database files, which can improve the buffer efficiency and reduce the page I/Os of the buffer manager.We compare PBM with three existing buffering algorithms, including LRU, LFU, and LeCaR, on two workloads, namely a regular workload and a shifting workload, which are composed of different access patterns.The results show that PBM outperforms the three competitors in terms of hit ratio and page I/Os.As a consequence, PBM achieves 1.66x, 2.03x, and 1.39x hit-ratio improvements compared to LRU, LFU, and LeCaR, respectively, on the regular workload.While on the shifting workload, PBM achieves 1.90x, 1.55x, and 1.49x higher hit ratios than LRU, LFU, and LeCaR, respectively.

Non-Volatile Memory (NVM) has emerged as an alternative to next-generation main memories. Although many tree indices have been proposed for NVM, they generally use B+-tree-like structures. To further improve the performance of NVM-aware indices, we consider integrating learned indexes into NVM. The challenges of such an integration are two fold: (1) existing NVM indices rely on small nodes to accelerate insertions with crash consistency, but learned indices use huge nodes to obtain a flat structure. (2) the node structure of learned indices is not NVM friendly, meaning that accessing a learned node will cause multiple NVM block misses. Thus, in this paper, we propose a new persistent learned index called PLIN. The novelty of PLIN lies in four aspects: an NVM-aware data placement strategy, locally unordered and globally ordered leaf nodes, a model copy mechanism, and a hierarchical insertion strategy. In addition, PLIN is proposed for the NVM-only architecture, which can support instant recovery. We also present optimistic concurrency control and fine-grained locking mechanisms to make PLIN scalable to concurrent requests. We conduct experiments on real persistent memory with various workloads and compare PLIN with APEX, PACtree, ROART, TLBtree, and Fast&Fair. The results show that PLIN achieves 2.08x higher insertion performance and 4.42x higher query performance than its competitors on average. Meanwhile, PLIN only needs ~30 μs to recover from a system crash.

Recently, Zoned Namespaces (ZNS) SSDs have emerged as a hot topic in both academics and industries. Compared to conventional SSDs, ZNS SSDs have the advantages of less overhead of garbage collection and lower over-provisioning cost. However, ZNS SSDs only accept sequential writes, and the zones inside ZNS SSDs need to be carefully managed to maximize the advantages of ZNS SSDs. Therefore, how to make data management systems adapt to ZNS SSDs is becoming a challenging issue. Current database systems, either SQL databases or NoSQL data stores, are mainly designed toward magnetic disks or traditional SSDs (without zoned namespaces). In this paper, we explore the challenges and research opportunities of revising index structures for ZNS SSDs and focus on the B+-tree and LSM-tree, which represent the index structures for SQL databases and key-value stores. After summarizing the features of ZNS SSDs, we discuss the key issues of adapting the B+-tree to ZNS SSDs and the challenges of revising the LSM-tree (Log-Structured Merge tree) for ZNS SSDs. Finally, we suggest some future research work on this topic.

Join selectivity estimation is a fundamental problem in query optimization, which aims to estimate the cardinality returned by a join query. While join selectivity estimation has been extensively studied in relational databases, there are few studies toward join selectivity estimation for cross-model joins in multi-model databases, such as relation-tree joins between a relational table and a tree-structured document like a BSON file in MongoDB. So far, due to the popularity of MongoDB, many applications use both MongoDB and MySQL to organize heterogeneous data. Thus, it is necessary to devise efficient approaches for processing relation-tree joins running on relational and tree models. In this paper, we present an effective and efficient approach to estimate the join selectivity for relation-tree joins, which consists of a value join estimation and a structural join estimation. In particular, we propose a two-level sampling method that samples the relational tuples and tree nodes at two levels. Then, we apply the discrete learning algorithm to the tree node samples to estimate the join value distribution of the tree nodes. With this mechanism, we can capture the correlation between relational tuples and tree nodes and improve the estimation accuracy. We conduct experiments on the DBLP dataset and compare our approach to existing solutions, and the results suggest the effectiveness and efficiency of our proposal.

Existing approaches based on the SGR (single granularity recognition) improve the recognition accuracy by expanding the dataset, learning more robust feature representation or locating and amplifying significant parts. These approaches often suffer from heavy computational costs, high model complexity or additional non-classsification losses. In this paper, we propose a feature fusion method based on multi-classification losses, which can classify different granularity features and make the network focus on more useful information. The complexity and computational costs of our model have only increased slightly, which can be ignored. The results of comparative experiments show the effectiveness of our proposal on Stanford Cars dataset.

LSM-tree is widely used as a write-optimized storage engine in many NoSQL systems. However, the periodical compaction operations in LSM-tree cost many I/O bandwidths and CPU resources of the local server, resulting in throughput drops of the system. To address this issue, this paper proposes a new compaction scheme based on the FaaS (Functions as a Service) architecture, which is called FaaS Compaction. It utilizes the elastic computing capability of FaaS and always pushes compactions to a FaaS cluster. The FaaS cluster will perform actual compaction operations, which will not affect the processing of the local server. Therefore, we can maintain stable performance even when periodical compactions are triggered. We also present a Parallel Slight Compaction method to solve the timeout problem caused by heavy compactions. We implement the FaaS Compaction based on TerarkDB and a real FaaS cluster and experimentally compare the FaaS Compaction with the RocksDB's local compaction scheme and the state-of-the-art offloading compaction policy. The results suggest the efficiency, stability, and elasticity of our proposal.

LSM-tree is widely used as a write-optimized storage engine in many key-value stores. However, the periodical compaction operations in LSM-tree cost many I/O bandwidths and CPU resources of the local server, resulting in throughput drops of the system. To address this issue, this paper proposes a new compaction scheme based on a FaaS (Functions as a Service) cluster, which is called FaaS Compaction. It utilizes the elastic computing capability of FaaS clusters and pushes compactions to a FaaS cluster. The FaaS cluster will perform actual compaction operations, which will not affect the processing of the local server. Therefore, we can maintain stable performance even when periodical compactions are triggered. We implement the FaaS Compaction and compare the FaaS Compaction with RocksDB and the state-of-the-art offloading compaction policy. The results suggest the efficiency and elasticity of our proposal.

With the rapid advance of Non-Volatile Memory (NVM), it has been a hot topic to improve traditional tree indices like B+-tree for NVM. However, due to the high cost of the writing operations on NVM, few existing tree indices can offer high performance for both read and write operations. For example, the WB-tree with unsorted leaf nodes is write-optimized but has poor search performance. To address this problem, in this paper, we propose a read/write-optimized tree index called TLBtree (Two-Layer B+-tree) for NVM. TLBtree consists of a read-optimized top layer and a write-optimized bottom layer. We notice that the top levels of a B+-tree are read frequently, while the bottom levels are written frequently. Motivated by such an observation, we propose to design a read-optimized top layer and a write-optimized layer for the TLBtree index. We offer several read optimizations to implement the top layer and employ write-optimized structures to organize the bottom layer. With this mechanism, we can alleviate the read and write tradeoff of the index on NVM. We conduct extensive experiments on a server with Intel Optane DC Persistent Memory and compare TLBtree with state-of-the-art NVM-based tree indices, including WB-tree, Fast&fair, and FPtree. The results show that TLBtree outperforms other indices in write-intensive workloads by up to 1.7x throughput and achieves comparable read-only performance with read-optimized indices.

This paper addresses a domain adaptive person re-identification network to learn representations across domains for a challenging task: exploiting more robust features from a labeled source domain dataset and an unlabeled target domain dataset. Although convolutional networks have achieved high accuracy in single-domain re-identification task, the performance of trained models slumped in a new dataset. Current methods focus on global features, but part-based features can offer more details of pedestrians, which was been proved in recent works. In this paper, we consider local features as part invariance to reduce the gap between domains and learn discriminative features for cross-domain person re-identification. A part invariance network is proposed which implements a memory to store features extracted in the target domain to reduce camera and neighborhood invariance to improve the adaptability of the model. Experiments conducted show that the invariance proposed in this paper outperforms the existing methods by a large margin.

In this paper, we demonstrate Post2Story, which aims to detect events and generate storylines on microblog posts. Post2Story has several new features: (1) It proposes to employ social influence to extract events from microblogs. (2) It presents a new Event Graph Convolutional Network (E-GCN) model to learn the latent relationships among events, which can help predict the story branch of an event and link events. (3) It offers a user-friendly interface to extract and visualize the development of events. After an introduction to the system architecture and key technologies of Post2Story, we demonstrate the functionalities of Post2Story on a real dataset.

Non-volatile memory (NVM) as a new type of storage technology has many advantages such as non-volatility, byte addressability, high storage-density, and low energy consumption.Meanwhile, NVM has some limitations, e.g., asymmetric read and write latency, limited write endurance, and high price.Therefore, at present, it is not realistic to completely replace DRAM with NVM in computer systems.A more feasible scheme is to adopt the hybrid memory architecture composed of NVM and DRAM.Following the assumption of hybrid memory architecture, this paper proposes an NVM-friendly sorting algorithm called NVMSorting.Particularly, we introduce a new concept called natural runs to improve the existing MONTRES algorithm and present the cost analysis of the algorithm in the hybrid memory architecture.In order to verify the performance of our proposal, we implement six existing sorting algorithms as baselines, including the MONTRES algorithm, and conduct comparative experiments on an unsorted dataset and a partially sorted dataset.The experimental results suggest the efficiency of NVMsorting in terms of execution time and NVM writes.Especially, on the partially sorted dataset, NVMSorting has 6.2% improvement on time performance and 5.7% reduction on NVM writes compared to MONTRES, and 13.0% performance improvement and 27.1% NVM-write reduction compared to the traditional merge sorting algorithm.

Bloom filter is an efficient technique to improve query performance in LSM-tree-based databases, such as RocksDB, HBase, and Cassandra.However, the original Bloom filter uses a fixed false positive rate (FPR), which makes it inefficient for mixed queries that involve both point and range queries.To solve this problem, in this paper, we present an improved Bloom filter called RoBF (Range-Query-Oriented Bloom Filter), which uses a mixture of Bloom filters and can process mixed queries on LSM-tree efficiently.We design an efficient algorithm for generating the solution based on the query distribution.We compare our proposal with the trie-based filter and find out that each has its own advantages for various scenarios.Therefore, we propose to use different filters with varied sizes for different levels on LSM-tree.Following this idea, we present an algorithm to generate specific filters with a specific size for different levels on LSM-tree to optimize the performance of mixed queries under limited memory space.We conduct comparative experiments and compare the proposed RoBF with various competitors, and the results show that RoBF can improve the performance of evaluating mixed queries by up to 6x to 30x, compared to the original Bloom filter in RocksDB.

This paper demonstrates a prototype system called Post2Event that aims to extract key events from microblogs.While many events are hidden in microblogs, people may only care about those critical events, which are named key events in Post2Event.Specially, we propose to model the topic-related significance of an event and integrate the influence with the temporal characteristics of the event to measure the event's importance.We briefly present the architecture and technical details of Post2Event.Then, we report the comparative results of Post2Event on a real dataset.Finally, we demonstrate the running process of the system.

Non-volatile memory (NVM) has emerged as an alternative of the next-generation memory due to its non-volatility, byte addressability, high storage-density, and low-energy consumption. However, NVM also has some limitations, e.g. asymmetric read and write latency. Therefore, at present, it is not realistic to completely replace DRAM with NVM in computer systems. A more feasible scheme is to adopt the hybrid memory architecture composed of NVM and DRAM. Following the assumption of hybrid memory architecture, in this paper, we propose an NVM-friendly sorting algorithm called NVMSorting. Particularly, we introduce a new concept called Natural Run to improve the existing MONTRES algorithm. Further, we apply the proposed NVMSorting to database join algorithms to improve the performance of the existing sort-merge join. To verify the performance of our proposal, we implement six existing sorting algorithms as baselines, including the MONTRES algorithm, and conduct comparative experiments on real Intel Optane DC persistent memory. The results show that NVMSorting outperforms other sorting algorithms in terms of execution time and NVM writes. In addition, the results of the join experiment show that the NVMSorting algorithm achieves the highest performance among all schemes. Especially, in the partially ordered data, the execution time of NVMSorting is 2.9%, 2.7%, and 4.2% less than MONTRES, external sort, and quick sort, respectively. Also, the amount of NVM writes of the NVMSorting is 26.1%, 43.6%, 96.2% less than MONTRES, external sort, and quick sort, respectively.

The advance of byte-addressable persistent memory (PM) makes it a hot topic to revisit traditional tree indices such as B+-tree and radix tree, and a few new persistent memory-friendly tree indices have been proposed. However, due to the special features of persistent memory compared to DRAM and the limitations of B+-tree-like indices, it is much harder to optimize both search and write performance for tree indices on persistent memory. As a result, most existing indices for persistent memory, e.g., WB-tree, proposed to improve write performance while sacrificing search performance. Aiming to optimize both write and search performance for tree indices on persistent memory, in this paper, we first propose a novel Two-Layer Architecture (TLA) for constructing tree indices on persistent memory. The key idea, of TLA is to organize the index with a search-optimized top layer and a write-optimized bottom layer, letting the top layer optimize search performance and the bottom layer improve write performance. By adopting efficient structures for the two layers, TLA can boost both write and search performance for tree indices on persistent memory. Following the TLA architecture, we present a new index called TLBtree (Two-Layer B+-tree) offering high search and write performance for persistent memory. Moreover, we develop a concurrent TLBtree to support non-blocking read operations in multi-core environment. We evaluate our proposals under a server equipped with real Intel Optane persistent memory. The results show that TLBtree outperforms the state-of-the-art tree indices, including WB-tree, Fast&Fair, and FPTree, in both search and write performance. Also, the concurrent TLBtree can achieve up to 3.7x speedup than its competitors under the multi-core environment.

Data-intensive systems are typical applications in the big data era. Most data-intensive systems employ LSM (Log-Structured Merge)-tree-based key-value stores, such as RocksDB and LevelDB. The LSM-tree has a multi-level data structure and can transform random writes into sequential ones by a special operation called compaction. However, compaction operations are costly and will increase the processing latency and lower the throughput of the LSM-tree. In this paper, we propose an adaptive lazy compaction method called ALC for the LSM-tree. The main idea of ALC is to delay compactions operations when the current workload is high, and to perform compactions when the system is low-loaded. With this mechanism, we can lower the impact of compactions on the system's performance and achieve high stability and low latency. Our preliminary experiments show that ALC outperforms RocksDB and the state-of-the-art method SILK in latency and throughput-stability.

In this paper, to improve the read performance of LSM-tree, we propose an enhanced LSM-tree called HotKey-LSM. The key idea of HotKey-LSM is to put hot keys and cold keys in two separated column families. Thus, when a hot key is not in the block cache, LSM-tree only needs to access a small hot-key LSM-tree to read the key. With this mechanism, most hot-key requests will be answered with low latency: if the hot key is in the block cache, we can return the memory address of the key; if it is not in the cache, we search the small hot-key LSM-tree. This differs from the traditional LSM-tree in that a hot-key request may traverse a large LSM-tree, which causes a high read latency. We implement HotKey-LSM on RocksDB and compare HotKey-LSM with the original RocksDB. The result in terms of QPS suggests the efficiency of our proposal.

Classification for big image data is an essential field of big data technology. Big image data classification has a wide range of applications, such as face recognition, object detection, and human posture recognition. Traditional methods for big image data classification usually ignore the close connection between different levels of features in neural networks. In this paper, we propose a new way for big image data classification that improves the previous method through the Multi-Level Feature Connection Network (MFC-Net). MFC-Net presents a multi-level feature connection module to mine more useful features for classification. Experiments show that MFC-Net significantly improves its baseline SSD (Single Shot Multibox Detector) and other advanced classification neural networks in the classic datasets Pascal VOC and Microsoft COCO.

Non-Volatile Memory (NVM) has emerged as an alternative to next-generation memories. Compared to the traditional DRAM, NVM offers data persistency and higher density. However, so far, NVM has higher accessing latency than DRAM. Therefore, to ensure the high performance of data accessing, we still need to consider using DRAM in memory architecture. This leads to the hybrid memory architecture involving DRAM and NVM. Some previous benchmark works have shown that such hybrid memory architecture is more efficient than NVM-only architecture. Due to NVM's unique properties, the traditional memory B+-tree becomes unsuitable for NVM because of its high cost of maintaining node orderliness and high space-filling feature. In this paper, we propose to optimize the Adaptive Radix Tree (ART) for the hybrid memory architecture and offer a new index called HART (Hybrid Adaptive Radix Tree). HART takes advantage of ART's deterministic structure to get good query performance. Meanwhile, we only selectively persist linked list to reduce NVM access cost. In particular, we exploit the compression path to improve the leaf node's space utilization, making the subtree shorter. We run a preliminary experiment on a server with Intel Optane DC Persistent Memory and compare HART with several NVM-aware indexes. The results suggest the efficiency of our proposal.

In this paper, we propose RotaryDS to provide fast storage service for massive data streams. RotaryDS uses a rotation storage model, which employs distributed data buckets to accept highly-arriving data streams. All data buckets have a state, i.e., they can be in the state of data idle waiting, data filling, write waiting, and data dumping. The state of a data bucket is changed according to the data operations. With the rotation storage model, we distribute massive data streams among multiple data buckets, thereby improving the write throughput of the storage system. We implement RotaryDS based on the rotation storage model and conduct preliminary experiments to compare it with MongoDB. The results suggest the efficiency of our proposal.

Non-volatile memory (NVM) as a new kind of future memory has several special properties such as non-volatility, read/write asymmetry, and byte address-ability. This makes it difficult to directly replace DRAM with NVM in the current memory hierarchy. Thus, a practical way is to construct a hybrid memory composed of both NVM and DRAM. Such hybrid memory architecture introduces many new challenges for existing algorithms. In this paper, we focus on the data placement issue in NVM-based hybrid memory systems, i.e., how to place the data on DRAM and NVM for a data-centric algorithm so that it can achieve high performance on hybrid memory. Particularly, we propose an optimal data placement model (ODP) to properly store data structures on DRAM and NVM during the execution of an algorithm. We present the theoretical proof to ODP to ensure the correctness of the model. To demonstrate the efficiency of ODP, we apply the ODP to two kinds of data-centric algorithms, namely sorting and database join. For sorting algorithms, we implement four ODP-based sorting algorithms, including Insertion Sort, Selection Sort, Heapsort, and Merge Sort. For join algorithms, we implement four ODP-based join strategies, including Nested Loops Join, Sort Join, Hash Join, and Virtual Partitioning Join. We conduct comparative experiments to evaluate the performance of the sorting/join algorithms. The results show that the ODP-based sorting/join strategies are much faster than the classical sorting/join algorithms that are not NVM-aware. In addition, the ODP-based implementation can reduce more NVM writes, showing that it is more NVM-friendly.

The performances of object detectors are highly impacted by the discrepancy between existing data sets and application scenarios, leading to the so-called domain shift problem. Previous works, based on Faster R-CNN, focus on aligning the image-level features and the region-level features. However, the Region Proposal Network (RPN), as a key module between the image-level and the region-level modules, still has the problem of domain shift that leads to inaccurate or even false detected results. To tackle this issue, we propose a new design, Domain-Invariant RPN (DIR), which adopts adversarial learning to eliminate the domain shift in RPN, and thereby, significantly improving the accuracy and robustness of bounding box proposals. Furthermore, we propose a Double-Consistency Regularization (DCR) to improve the overall feature alignment. Extensive experiments show that our approach outperforms state-of-the-art methods.

With the development of Web 2.0, social media such as Twitter and Sina Weibo have become an essential platform for disseminating hot events. Simultaneously, due to the free policy of microblogging services, users can post user-generated content freely on microblogging platforms. Accordingly, more and more hot events on microblogging platforms have been labeled as spammers. Spammers will not only hurt the healthy development of social media but also introduce many economic and social problems. Therefore, the government and enterprises must distinguish whether a hot event on microblogging platforms is a spammer or is a naturally-developing event. In this paper, we focus on the hot event list on Sina Weibo and collect the relevant microblogs of each hot event to study the detecting methods of spammers. Notably, we develop an integral feature set consisting of user profile, user behavior, and user relationships to reflect various factors affecting the detection of spammers. Then, we employ typical machine learning methods to conduct extensive experiments on detecting spammers. We use a real data set crawled from the most prominent Chinese microblogging platform, Sina Weibo, and evaluate the performance of 10 machine learning models with five sampling methods. The results in terms of various metrics show that the Random Forest model and the over-sampling method achieve the best accuracy in detecting spammers and non-spammers.

Flash memory based solid state drives (SSDs) have emerged as a new alternative to replace magnetic disks due to their high performance and low power consumption. However, random writes on SSDs are much slower than SSD reads. Therefore, traditional index structures, which are designed based on the symmetrical I/O property of magnetic disks, cannot completely exert the high performance of SSDs. In this paper, we propose an SSD-optimized linear hashing index called Self-Adaptive Linear Hashing (SALhashing) to reduce small random-writes to SSDs that are caused by index operations. The contributions of our work are manifold. First, we propose to organize the buckets of a linear hashing index into groups and sets to facilitate coarse-grained writes and adaptivity to access patterns. A group consisting of a fixed number of buckets is proposed to transform small random writes to buckets into coarsegrained writes and in turn improve write performance of the index. A set consists of a number of groups, and we propose to employ different split strategies for each set. With this mechanism, SAL-hashing is able to adapt to the changes of access patterns. Second, we attach a log region to each set, and amortize the cost of reads and writes by committing updates to the log region in batch. Third, in order to reduce search cost, each log region is equipped with Bloom filters to index update logs. We devise a cost-based online algorithm to adaptively merge the log region with the corresponding set when the set becomes search-intensive. Fourth, we propose a new technique called virtual split to optimize the search performance of SAL-hashing. Finally, we propose a new scheme for the management of the log buffer. We conduct extensive experiments on real SSDs. The results suggest that our proposal is self-adaptive according to the change of access patterns, and outperforms several competitors under various workloads.

On-Line Analytical Processing (OLAP) enables powerful analytics by quickly computing aggregate values of numerical measures over multiple hierarchical dimensions for massive datasets. However, many types of source data, e.g., from GPS, sensors, and other measurement devices, are intrinsically inaccurate (imprecise and/or uncertain) and thus OLAP cannot be readily applied. In this paper, we address the resulting data veracityproblem in OLAP by proposing the concept of probabilistic data cubes. Such a cube is comprised of a set of probabilistic cuboids which summarize the aggregated values in the form of probability mass functions (pmfs in short) and thus offer insights into the underlying data quality and enable confidence-aware query evaluation and analysis. However, the probabilistic nature of data poses computational challenges, since a probabilistic database can have exponential number of possible worlds under the possible world semantics. Even worse, it is hard to share computations among different cuboids, as aggregation functions that are distributive for traditional data cubes, e.g., SUM, become holistic in probabilistic settings. In this paper, we propose a complete set of techniques for probabilistic data cubes, from cuboid aggregation, over cube materialization, to query evaluation. We study two types of aggregation: convolution and sketch-based, which take polynomial time complexities for aggregation and jointly enable efficient query processing. Also, our proposal is versatile in terms of: 1) its capability of supporting common aggregation functions, i.e., SUM, COUNT, MAX, and AVG; 2) its adaptivity to different materialization strategies, e.g., full versus partial materialization, with support of our devised cost models and parallelization framework; 3) its coverage of common OLAP operations, i.e., probabilistic slicing and dicing queries. Extensive experiments over real and synthetic datasets show that our techniques are effective and scalable.

There are a great number of complex relations among different characters in an opera. Retrieving such relations is crucial for performers and audience to accurately understand the features and behaviour of roles. Aiming to automatically extract relations among characters in an opera, in this paper we propose an effective method that can extract character relations from opera scripts. Firstly, we construct a uniform reasoning framework for opera scripts. Based on this model, we propose a deep syntax-parsing method to detect character relations from opera scripts. After that, we propose a new deep learning approach called SL-Bi-LSTM-CRF to extract the objects involved in character relations. The proposed SL-Bi-LSTM-CRF algorithm is a sentence-level relation extraction algorithm based on the Bi-directional LSTM with a CRF layer. With this mechanism, we are able to get a detailed description for character relations. We conduct experiments on a real dataset of opera scripts. The experimental results in terms of precision, recall, and F-score suggest the effectiveness of our proposal.

Hadoop, especially Hadoop 2.0, has been a dominant framework for real-time big data processing. However, Hadoop is not optimized for energy efficiency. Aiming to solve this problem, in this paper, we propose a new framework to improve the energy efficiency of Hadoop 2.0. We focus on the resource manager in Hadoop 2.0, namely YARN, and propose energy-efficient task scheduling mechanisms on YARN. Particularly, we focus on CPU-intensive streaming jobs and classify streaming jobs into two types, namely batch streaming jobs (i.e., a set of jobs are submitted simultaneously) and online streaming jobs (i.e., jobs are continuously submitted one by one). We devise different energy-efficient task scheduling algorithms for each kind of streaming jobs. Specially, we first propose to abstractly model performance and energy consumption by considering the characteristics of tasks as well as the computational resources in YARN. Based on this model, we study the energy efficiency of streaming tasks which consist of the performance model and energy consumption model of task. We propose two key principles for improving energy efficiency: 1) CPU usage aware task allocation, partitions tasks to NMs based on the task characteristic in term of CPU usage; and 2) resource efficient task allocation, reduce idle resource. Then, we propose a D-based binning algorithm for the batch task scheduling and K-based binning algorithm for the online task scheduling that can adapt to continuously arriving tasks. We conduct extensive experiments on a real Hadoop 2.0 cluster and use two kinds of workloads to evaluate the performance and energy efficiency of our proposal. Compared with Storm (the streaming data processing tool in Hadoop 2.0) and other approaches including TAPA and DVFS-MR, our proposal is more energy efficient. The batch task scheduling algorithm reduces up to 10 percent of energy consumption and keeps comparable performance. In addition, the online task scheduling algorithm reduces up to 7 percent over the existing algorithms.

With the development of networked radars and wireless communication technologies, electromagnetic data is becoming a new type of big data.Compared with other kinds of big data, such as Internet big data, financial big data, and healthcare big data, electromagnetic big data has some special properties.For example, they usually contain rich and varying labels that describe the features of electromagnetic space.However, existing big-data benchmark tools cannot support the generation and test of electromagnetic big data.In this paper, we aim at providing a simulation-based testbed for electromagnetic big data.The testbed, named STEM, can simulate real-world electromagnetic big data.It supports generating real-time electromagnetic data streams with varying labels.In addition, it is reusable, reconfigurable, and flexible for users to generate workloads for different scenarios.After a brief introduction on the architecture of STEM, we present the implemental details of STEM.Then, we present a case study as well as performance evaluation to demonstrate the usability and flexibility of STEM.

Company acquisition relation reflects a company's development intent and competitive strategies, which is an important type of enterprise competitive intelligence. In the traditional environment, the acquisition of competitive intelligence mainly relies on newspapers, internal reports, and so on, but the rapid development of the Web introduces a new way to extract company acquisition relation. In this paper, the authors study the problem of extracting company acquisition relation from huge amounts of Web pages, and propose a novel algorithm for company acquisition relation extraction. The authors' algorithm considers the tense feature of Web content and classification technology of semantic strength when extracting company acquisition relation from Web pages. It first determines the tense of each sentence in a Web page, which is then applied in sentences classification so as to evaluate the semantic strength of the candidate sentences in describing company acquisition relation. After that, the authors rank the candidate acquisition relations and return the top-k company acquisition relation. They run experiments on 6144 pages crawled through Google, and measure the performance of their algorithm under different metrics. The experimental results show that the algorithm is effective in determining the tense of sentences as well as the company acquisition relation.

change memory (PCM) as a newly developed storage medium has many attractive properties such as non-volatility, byte addressability, high density and low energy consumption. Thus, PCM can be used to build non-volatile main memory databases. However, PCM's long write latency and high write energy bring challenges to PCM-based memory systems. In this paper, we propose an improvement over the B+-tree for PCM. Particularly, we consider the read/write tendency of leaf nodes. For write-intensive leaf nodes, we use an overflow-node technique to reduce PCM writes, while for read-intensive ones, we adjust the tree structure to remove overflow nodes to improve read performance. Our experimental results suggest that our proposal outperforms the traditional B+-tree and the overflow B+-tree.

On-Line Analytical Processing (OLAP) enables powerful analytics by quickly computing aggregate values of numerical measures over multiple hierarchical dimensions for massive datasets. However, many types of source data, e.g., from GPS, sensors, and other measurement devices, are intrinsically inaccurate (imprecise and/or uncertain) and thus OLAP cannot be readily applied. In this paper, we address the resulting data veracity problem in OLAP by proposing the concept of probabilistic data cubes. Such a cube is comprised of a set of probabilistic cuboids which summarize the aggregated values in the form of probability mass functions (pmfs in short) and thus offer insights into the underlying data quality and enable confidence-aware query evaluation and analysis. However, the probabilistic nature of data poses computational challenges as even simple operations are #P-hard under the possible world semantics. Even worse, it is hard to share computations among different cuboids, as aggregation functions that are distributive for traditional data cubes, e.g., SUM and COUNT, become holistic in probabilistic settings. In this paper, we propose a complete set of techniques for probabilistic data cubes, from cuboid aggregation, over cube materialization, to query evaluation. For aggregation, we focus on how to maximize the sharing of computation among cells and cuboids. We present two aggregation methods: convolution and sketch-based. The two methods scale down the time complexities of building a probabilistic cuboid to polynomial and linear, respectively. Each of the two supports both full and partial data cube materialization. Then, we devise a cost model which guides the aggregation methods to be deployed and combined during the cube materialization. We further provide algorithms for probabilistic slicing and dicing queries on the data cube. Extensive experiments over real and synthetic datasets are conducted to show that the techniques are effective and scalable.

We study safe zones for service subscriptions in a volunteered geographic service setting, covering the concepts, properties, and algorithms needed for the use of weighted imprecise Voronoi cells as safe zones. Empirical performance studies on both synthetic and real datasets offer insights into the efficiency and scalability of our proposal. For a comprehensive coverage, see the full version of the paper.

Flash memory based solid state drives (SSDs) have emerged as a new alternative to replace magnetic disks due to their high performance and low power consumption. However, random writes on SSDs are much slower than SSD reads. Therefore, traditional index structures, which are designed based on the symmetrical I/O property of magnetic disks, cannot completely exert the high performance of SSDs. In this paper, we propose an SSD-optimized linear hashing index called Self-Adaptive Linear Hashing (SAL-Hashing) to reduce small random writes to SSDs that are caused by index operations. The contributions of our work are manifold. First, we propose to organize buckets into groups and sets to facilitate coarse-grained writes and lazy-split so as to avoid intermediate writes on the hash structure. A group consists of a fixed number of buckets and a set consists of a number of groups. Second, we attach a log region to each set, and amortize the cost of reads and writes by committing updates to the log region in batch. Third, in order to reduce search cost, each log region is equipped with Bloom filters to index update logs. We devise a cost-based online algorithm to adaptively merge the log region with the corresponding set when the set becomes search-intensive. Finally, in order to exploit the internal package-level parallelisms of SSDs, we apply coarse-grained writes for merging or split operations to achieve a high bandwidth. Our experimental results suggest that our proposal is self-adaptive according to the change of access patterns, and outperforms several competitors under various workloads on two commodity SSDs.

We provide techniques that enable a scalable so-called Volunteered Geographic Services system. This system targets the increasing populations of online mobile users, e.g., smartphone users, enabling such users to provide location-based services to each other, thus enabling citizen reporter or citizen as a sensor scenarios. More specifically, the system allows users to register as service volunteers, or micro-service providers, by accepting service descriptions and periodically updated locations from such volunteers; and the system allows users to subscribe to notifications of available, nearby relevant services by accepting subscriptions, formalized as continuous queries, that take service preferences and user locations as arguments and return relevant services. Services are ranked according to their relevance and distance to a query, and the highest ranked services are returned. The key challenge addressed is that of scalably providing up-to-date results to queries when the query locations change continuously. This is achieved by the proposal of a new so-called safe-zone model. With safe zones, query results are accompanied by safe zones with the property that a query result remains the same for all locations in its safe zone. Then, query users need only notify the system when they exit their current safe zone. Existing safe-zone models fall short in the papers setting. The new model is enabled by (i) weighted and (ii) set weighted imprecise Voronoi cells. The paper covers underlying concepts, properties, and algorithms, and it covers applications in VGS tracking and presents findings of empirical performance studies.

The rapid development of Internet of Things (IoT) enables people to track objects by deploying multiple sensors, e.g., to track people in indoor spaces using RFID sensors. Multi-sensor object-tracking data are usually produced as records, which are thereby organized into small files and written to servers. However, the small-size property and high arriving-rate of multi-sensor object-tracking data will result in poor I/O performance of file systems such as HDFS. In this paper, we propose the first read/write-optimized solution for storing multi-sensor object-tracking data on HDFS. In particular, we exploit a distributed caching mechanism and a parallel file-merging policy to improve the I/O performance of HDFS. With our design, object-tracking data are first cached by a Distributed Memory File System (DMFS) on top of HDFS. These data are further merged into large files, which are then flushed to HDFS in parallel. We demonstrate that this mechanism is able to improve the write throughput of HDFS and outperforms existing centralized-cache-based approaches. In addition, in order to improve the search performance of object queries over multi-sensor object-tracking data, we propose a Sensor-Dependence Graph (SDG) to model sensor dependence and further present an SDG-based algorithm to efficiently cluster sensors. The object-tracking data from the sensors in the same cluster are merged into the same large files, which can reduce file scans during query processing and therefore improve search performance. We conduct extensive experiments to evaluate the performance of our proposal. The results suggest the efficiency of our proposal with respect to disk-write throughput, memory-write throughput, search performance, and sensor clustering.

Phase change memory (PCM) has become one of the most promising storage media particularly for memory systems, due to its byte addressability, high access speed, and low energy consumption. In addition, hybrid memory systems involving both PCM and DRAM can utilize the merits of both media and overcome some typical drawbacks of PCM such as high write latency and limited lifecycle. In this paper, we present a novel page replacement algorithm called APP-LRU (Access-Pattern-prediction-based LRU) for PCM/DRAM-based hybrid memory systems. APP-LRU aims to reduce writes to PCM while maintaining stable time performance. Particularly, we detect read/write intensity for each page in the memory, and put read-intensive pages into PCM while placing write-intensive pages in DRAM. We conduct trace-driven experiments on six synthetic traces and one real OLTP trace. The results show that our proposal is able to reduce up to 5 times of migrations more than its competitors.

Indoor moving objects management has been a research focus in recent years. In order to get the trajectory data of indoor moving objects, people have to deploy a lot of positioning equipment, such as RFID readers and tags, which takes lots of money, time, and other costs. In addition, it is a very complex and costly process to construct different environment settings for various indoor applications. Aiming to provide experimental trajectory data for various indoor operations and mining algorithms, in this paper we present a flexible tool to generate trajectories for indoor moving objects, which is named IndoorSTG (Indoor Spatiotemporal Trajectory Generator). IndoorSTG can simulate different indoor environments using various elements including rooms, doors, corridors, stairs, elevators, and virtual positioning devices such as RFID or Bluetooth readers. Meanwhile, it can generate semantic-based trajectories for indoor moving objects in a specific indoor space. After an overview of the general features of IndoorSTG, we discuss the architecture and implementation of IndoorSTG. And finally, a case study of IndoorSTG's demonstration is presented.

Most Web pages contain temporal information, which can be utilized by search engines to improve searching performance for users. However, traditional search engines have little support in processing temporal-textual Web queries. Aiming at solving this problem, in this paper we present and implement a prototype system for time-sensitive queries, which is called TASE (Time-Aware Search Engine). TASE extracts both the explicit and implicit temporal expressions for each Web page, and calculates the relevant score between the Web page and each temporal expression, and then re-rank search results based on the temporal-textual relevance between Web pages and the queries. It is demonstrated that TASE can improve the effectiveness of temporal-textual Web queries.

In content-based image retrieval (CBIR), color is the most intuitive image features and it is widely used. But The current color feature can describe the semantics of the whole image effectively, but does not reflect characteristics of the color salience objects in an image. For the purpose of giving prominence to the color characteristics of salience objects, this paper proposes a new color feature descriptor based on color visual perception model. By analyzing the human color visual perception process, the color perception model is proposed at first. This model integrates the intensity, the color contrast and self saliency, sparsity and centricity saliency to describe human color visual perception of the image. Then, the new color feature descriptor is calculated by weighting the significant bit-plane histograms with color perception map. Finally, similarity measure is presented for the new color feature. Experiment results show that the proposed color feature is more accurate and efficient in retrieving images with user-interested color objects. Compared with the other three retrieval methods, the proposed technique improves the retrieval accuracy effectively.

Flash disks have become a popular alternative for magnetic disks, due to their fast I/O speed and other features such as small-size, shock-resistance, energy-efficient and non-volatile. However, flash disks also have characteristics of out-of-place update and asymmetric I/O latencies for read, write, and erase operations, which introduce new challenges into the indexing mechanism for flash disks. Traditional index structures do not take the flash I/O characteristics into account and, therefore, will cause poor performance. To address this problem, we present a new hybrid index structure for flash disks in this paper, which is called HashTree. HashTree aims at getting better update performance while keeping relatively high search efficiency. The HashTree uses a hash-based index to split the indexed records into several buckets, and then we develop a new tree structure named FTree to organize the records in each bucket. Compared with the previous tree-based indexes, our policy can reduce the costs of maintaining the hierarchical tree structure. We also introduce a tuning mechanism into the HashTree so that we can obtain appropriate trade-off between search performance and update performance. We conducted an experiment on a commercial SSD to evaluate the performance of HashTree. Compared with its competitors such as BFTL and FD-tree, our experimental results show that HashTree performs best in terms of both update performance and overall performance.

In this paper, we present a flexible simulation environment for the performance evaluation of flash-aware algorithms, which is called Flash-DBSim. The main purpose of Flash-DBSim is to provide a configurable virtual flash disk for upper systems, such as file system and DBMS, so that the algorithms in those systems can be easily evaluated on different types of flash disks. Moreover, it also offers a prototyping environment for those algorithms inside flash disk, e.g. the algorithms for garbage collection or wear-leveling. After an overview of the general features of Flash-DBSim, we discuss the architecture of Flash-DBSim. And finally, a case study of Flash-DBSim's demonstration is presented.

Flash disks are becoming an important alternative to conventional magnetic disks. Although accessed through the same interface by applications, flash disks have some distinguished characteristics that make it necessary to reconsider the design of the software to leverage their performance potential. This paper addresses this problem at the buffer management layer of database systems and proposes a flash-aware replacement policy that significantly improves and outperforms one of the previous proposals in this area.

In this paper, we present a CASE tool supporting conceptual modeling for moving objects database applications, which is called STXER (spatio-temporal extended entity-relational model). The main purpose of STXER is to support the database design for moving objects applications on typical object-relational DBMS. The STXER tool enhances the traditional ER model with moving characteristics. It can express rich spatiotemporal semantics for moving objects applications. After an overview of the general features of STXER, we discuss the architecture of STXER. And finally, a case study of STXER's demonstration is presented.