Vector approximate nearest neighbor search (ANNS) underpins search engines, recommendation systems, and advertising services. Recent advances in ANNS indexes make CPU a cost-effective choice for serving million-scale, in-memory vector search, yet per-core throughput remains constrained by memory access latency of vector reading and the compute intensity of distance evaluations in production deployments. With the growing scale of the business and advances in hardware, modern CCD-based multi-core CPUs have been widely deployed for high throughput in our services. However, we find that simply increasing core counts does not yield optimal performance scaling. To improve the efficiency of more cores from the CCD-based architecture, we analyze the distributions of real-world requests in our production environments. We observe high access locality in vector search in our online services and low cache utilization, resulting from overlooking the multi-chiplet nature of CCD based CPUs. Hence, we propose a workload- and hardware-aware thread orchestration framework at CCD-level that (i) provides a uniform interface for both inter-query parallel HNSW search and intra-query parallel IVF search, (ii) achieves cache-friendly and workload-adaptive mapping of task dispatching, and (iii) employs CCD-aware task stealing to address load imbalance. Applied to real production workloads from search, recommendation, and advertising services of Xiaohongshu (RedNote), our approach delivers up to 3.7x higher throughput and 30-90% reductions in P50 and P999 latency. In detail, compared with the original framework, the cache-miss ratio decreases by 6-30%, and the total CPU stall is reduced by 20-80%.

Many cloud databases provide fine-grained regular snapshots and sparsely deleted snapshots based on importance, and dynamically maintain large-scale snapshots to ensure data security and mine the value of cold data. However, in existing snapshot technologies, the write amplification feature of Copy-on-Write (CoW) introduces additional expensive I/O operations in a cloud environment. In Redirect-on-Write (RoW), the modified data blocks are scattered among the snapshots, resulting in a dependency between the snapshots, which seriously affects the recovery performance. In this paper, we observed that access to snapshots has the characteristics of locality and continuity. We therefore propose an efficient Batch-Based Snapshot index, called BBS, which batches snapshot indexes according to database workload and access behavior of snapshots. Specifically, we use two key techniques: Shared-Subtrees Indexing and Batch-Based Dividing, to perform split dependency of the snapshot index. The snapshot index dependency chain is divided into batches, and there is no dependency on snapshot indexes between batches. In-batch snapshot indexes reduce memory overhead by sharing subtrees. The index can directly locate data blocks instead of iterative traversal. At the same time, the design of the snapshot index deletion method is adapted to the snapshot sparse deletion model. We have implemented a working system in Ceph. Evaluation results on datasets demonstrate that, compared with existing techniques, BBS can effectively balance the overhead between index memory capacity and recovery time.

With the rapid development of storage and network technology, emerging high-performance hardware is being widely applied to the distributed storage cluster. However, existing distributed storage systems employing multi-layer abstractions to provide table data services result in leaving high-speed hardware under-exploited. In this paper, we propose TEngine, a native distributed table storage engine designed for NVMe SSD and RDMA. The key is that TEngine removes the file abstraction to construct table structures on the device directly. For metadata service, TEngine designs a decoupled single metadata server, reducing distributed coordination, easing the burden on the metadata node, and enabling localized data node access. For data service, TEngine optimizes the parallel processing capability of NVMe devices by integrating upper-level multi-thread parallel operations with lower-level NVMe devices' parallel I/O processing. Moreover, TEngine introduces a periodic pull-based data synchronization approach to transform data pushing into periodic data pulling, which offloads the synchronization burden from the leader to the followers. The experimental results show that TEngine outperforms state-of-the-art distributed storage systems using the same hardware environment.

With the development of AI generated content and LLM (Large Language Model), demands of vector management have brought prosperity to vector databases. However, the status that vectors cannot be retrieved before being indexed, harms timeliness of vector databases. Updating indexes immediately when adding new vectors, reduces throughput of storage. Due to this contradiction, when facing streaming data, using vector database solely in vector services cannot have it both ways: real-time searches and high-throughput storage. This paper proposes a vector buffer engine, Neos. It is designed for real-time unindexed-vector searches on streaming input and buffering vectors with high throughput before loading them into vector databases. On one hand, we build a lightweight storage on raw NVMe device and liberate throughput from indexes, to maximize storage performance. On the other hand, we realize direct NVMe-GPUs 110 stack and a CPU-GPU heterogeneous task architecture for low-latency unindexed-vector searches on streaming data. Experiments show that our approach performs with 1.5x to 3.4x bandwidth, as low as 20% latency compared to existing 110 stacks, and up to orders-of-magnitude higher vector storage throughput under concurrent RIW workloads. Further, N eos can handle real-time unindexed - vector searches with millisecond-level latency on streaming input, a capability that current vector systems lack.

The increasing capacity and reducing cost of the main memory made in-memory data analytics systems widely deployed as they could provide higher throughput and lower latency. Since the data resides in memory, computational throughput becomes a crucial factor in the performance of these systems rather than disk accesses. Single instruction multiple data (SIMD) is an effective mechanism to improve computational performance, which has been well studied to accelerate data analytics systems. However, the state-of-the-art methods focus on using SIMD more efficiently while neglecting scalar execution units.In this paper, we present the hybrid execution framework (HEF) to co-utilize SIMD and scalar execution units for the data analytics workload. We also extend the concept of pack to eliminate the data dependency between adjacent instructions, achieving shorter instruction execution intervals. Experimental results show that the hybrid execution achieves up to 2.38× and 1.45× better performance compared with the purely scalar and SIMD implementation on the star schema benchmark (SSB) queries, respectively. Besides, HEF performs better than the state-of-the-art system Voila for a majority of queries in SSB under all data scales.

Although real-time analytics on the up-to-date dataset has become an emerging demand, many big data systems are still designed for offline analytics. Particularly, for critical applications like Fintech,transactional data ingestionensures a timely, always-correct, and scalable dataset. To carry out append-only ingestion, existing OLTP/HTAP systems are based on strict transactions with imperfect scalability, while NoSQL-like systems support scalable but relaxed transactions. How to ensure essential transactional guarantees without harming scalability seems to be a non-trivial issue. This paper proposesKarstto bring transactional data ingestion for existing offline analytics. We notice that blockingtwo-phase commit(2PC) to resolve transactional data ingestion is a performance killer for the partitioned analytical systems. Karst introduces a scalable protocol calledmetadata-oriented commit(MOC) that converts each distributed transaction into multiple partial transactions to avoid 2PC. Moreover, to ingest massive data into plenty of partitions, Karst also employs lazy persistence, lightweight logging, and optimized data traffic. In experiments, Karst could achieve up to about 2x$\sim$10x performance over relevant systems and also shows remarkable scalability.

Distributed in-memory computing systems have dramatic performance improvement over traditional disk-based systems, which makes them widely used in large-scale data processing applications. Unfortunately, uneven and unpredictable data distributions caused by data skew have a significant impact on the performance. In Spark, when data skew happens, some tasks will process much more data than other tasks and become the performance bottleneck. The traditional approaches to handling data skew are based on sampling and repartitioning, which incur additional overhead. In this paper, we divide data skew in distributed data processing systems into intra-node and inter-node skew. Based on data stealing, we proposed DS2to handle both intra-node and inter-node data skew. It aims to improve the performance under data skew, without involving additional overhead. DS2first balances the skewed data distribution in the local and then handles the inter-node skew by RDMA during execution. It achieves up to 2.96× speedup on the aggregation operator and 2.81× speedup on the join operator.

Most database systems rely on complex multi-layer and compatibility-oriented storage stacks, which results in sub-optimal database management system (DBMS) performance and significant write amplification. A heavy storage stack can be tolerated in the slow disk era because its storage overhead is completely overlapped by hardware delay. However, with advances in storage technologies, emerging NVMe devices have reached the same level of latency as software, which in turn has caused the storage stack to become a new bottleneck. On the other hand, NVMe devices not only improve I/O efficiency but also introduce distinctive hardware features that require software modifications to take advantage of themTo fully exploit the hardware potential of NVMe devices, we propose a lightweight native storage stack called Lightstack to minimize the software overhead. The core of Lightstack is an efficient table storage engine, LATTE, which abstracts the essential data service of the database's 2D table. LATTE is designed from the ground up to use NVMe devices efficiently. It directly accesses NVMe devices to reduce single I/O latency and utilizes a parallel scheduling strategy to leverage multiple deep I/O queues and CPU cores. Besides, undo logging on heterogeneous storage is proposed to mitigate the write amplification further. We also implement a working prototype and evaluate it with standard benchmarks on the Intel Optane DC P4800X NVMe SSD and the DC P3608 Series NVMe SSD. Experimental results show that LATTE has up to 3.6-6.5× the throughput of MySQL's InnoDB and MyRocks engines, with latency as low as 28% in the same hardware environment.

To fully utilize the resources of a main memory database cluster, we additionally take the independent parallelism into account to parallelize multiple pipelines of one query. However, scheduling resources to multiple pipelines is an intractable problem. Traditional static approaches to this problem may lead to a serious waste of resources and suboptimal execution order of pipelines, because it is hard to predict the actual data distribution and fluctuating workloads at compile time. In response, we propose a dynamic scheduling algorithm, List with Filling and Preemption (LFPS), based on two novel techniques. (1) Adaptive filling improves resource utilization by issuing more extra pipelines to adaptively fill idle resource “holes” during execution. (2) Rank-based preemption strictly guarantees scheduling the pipelines on the critical path first at run time. Interestingly, the latter facilitates the former filling idle “holes” with best efforts to finish multiple pipelines as soon as possible. We implement LFPS in our prototype database system. Under the workloads of TPC-H, experiments show our work improves the finish time of parallelizable pipelines from one query up to 2.5X than a static approach and 2.1X than a serialized execution.

SIMD is an instruction set in mainstream processors, which provides the data level parallelism to accelerate the performance of applications. However, its advantages diminish when applications suffer from heavy cache misses. To eliminate cache misses in SIMD vectorization, we present interleaved multi-vectorizing (IMV) in this paper. It interleaves multiple execution instances of vectorized code to hide memory access latency with more computation. We also propose residual vectorized states to solve the control flow divergence in vectorization. IMV can make full use of the data parallelism in SIMD and the memory level parallelism through prefetching. It reduces cache misses, branch misses and computation overhead to significantly speed up the performance of pointer-chasing applications, and it can be applied to executing entire query pipelines. As experimental results show, IMV achieves up to 4.23X and 3.17X better performance compared with the pure scalar implementation and the pure SIMD vectorization, respectively.

To fully use the advanced resources of a main memory database cluster, we take independent parallelism into account to parallelize multiple pipelines of one query. However, scheduling resources to multiple pipelines is an intractable problem. Traditional static approaches to this problem may lead to a serious waste of resources and suboptimal execution order of pipelines, because it is hard to predict the actual data distribution and fluctuating workloads at compile time. In response, we propose a dynamic scheduling algorithm, List with Filling and Preemption (LFPS), based on two techniques. (1) Adaptive filling improves resource utilization by issuing more extra pipelines to adaptively fill idle resource "holes" during execution. (2) Cost-based preemption strictly guarantees scheduling the pipelines on a critical path first at run time. We implement LFPS in our prototype database system. Under the workloads of TPC-H, experiments show our work improves the finish time of parallelizable pipelines from one query up to 2.3X than a static approach and 1.7X than a serialized execution.

Hardware techniques and environments underwent significant transformations in the field of information technology, represented by high-performance processors and hardware accelerators characterized by abundant heterogeneous parallelism, nonvolatile memory with hybrid storage hierarchies, and RDMA-enabled high-speed network. Recent hardware trends in these areas deeply affect data management and analysis applications. In this paper, we first introduce the development trend of the new hardware in computation, storage, and network dimensions. Then, the related research techniques which affect the upper data management system design are reviewed. Finally, challenges and opportunities are addressed for the key technologies of data management and analysis in new hardware environments.

With the development and popularization of cloud computing, more and more services and applications are migrated to cloud for the sake of low cost, high availability and excellent performance. As the foundation of cloud computing, virtualization technology integrates and reallocates the computing capability, storage and network resource fairly among virtual machines and provides a full-featured, isolated and reliable hardware environment for various operating systems. Owe to the virtualization technology, computing capability of virtual machines has achieved fantastic performance, some even achieve near native speed. However, low I/O performance is still a bottleneck, especially in I/O intensive applications. The leading causes include redundant data copy and frequent VM exits. Focusing on network I/O optimization, we design and implement virtsocket, a new network socket library in virtualization scenario which utilizes shared memory for data transmission. A ring buffer data structure stores I/O requests of virtual machine which is triggered to issue all requests with only one hypercall according to scheduler. Data referred in the I/O requests is read directly from virtual machine memory by host machine kernel module with interfaces provided by modified hypervisor. Experimental results show that throughput is improved by hundreds of times when compared with original virtualization scenario, and the latency also achieves a remarkable reduction. Both throughput and latency performance exceed existing para-virtualization solutions.