Query optimization is essential for efficient SQL query execution in DBMS, and remains attractive over time due to the growth of data volumes and advances in hardware. Existing traditional optimizers struggle with the cumbersome hand-tuning required for complex workloads, and the learning-based methods face limitations in ensuring generalization. With the great success of Large Language Model (LLM) across diverse downstream tasks, this paper explores how LLMs can be incorporated to enhance the generalization of learned optimizers. Though promising, such an incorporation still presents challenges, mainly including high model inference latency, and the substantial fine-tuning cost and suboptimal performance due to inherent discrepancy between the token sequences in LLM and structured SQL execution plans with rich numerical features. In this paper, we focus on recurring queries in offline optimization to alleviate the issue of high inference latency, and propose \textbf{LLM4Hint} that leverages moderate-sized backbone LLMs to recommend query optimization hints. LLM4Hint achieves the goals through: (i) integrating a lightweight model to produce a soft prompt, which captures the data distribution in DBMS and the SQL predicates to provide sufficient optimization features while simultaneously reducing the context length fed to the LLM, (ii) devising a query rewriting strategy using a larger commercial LLM, so as to simplify SQL semantics for the backbone LLM and reduce fine-tuning costs, and (iii) introducing an explicit matching prompt to facilitate alignment between the LLM and the lightweight model, which can accelerate convergence of the combined model. Experiments show that LLM4Hint, by leveraging the LLM's stronger capability to understand the query statement, can outperform the state-of-the-art learned optimizers in terms of both effectiveness and generalization.

Learned plan-selection optimizers combine the conventional and learned approaches by generating diverse candidate plans through multiple optimizers and selecting the best via value models. However, these eagerly-generated plans incur high optimization overhead, as they require multiple invocations of the native optimizer. In this article, we propose MoEPlan, which learns a routing policy to select top- \( k \) experts (different optimizers) via query embedding and learnable parameters, avoiding pre-generation of candidate plans. Our approach integrates two optimization strategies: (1) a virtual ideal expert to guide the best plan selection through learned plan similarities, and (2) a query-irrelevant expert sampling strategy to balance the training cost and effectiveness of selected plans in the first round of expert selection. Furthermore, we design a two-phase training process: the first phase pre-trains the model with complete expert feedback, while the second phase filters the full expert pool to yield a promising subset and refines the selection to pinpoint the optimal expert. Experimental studies show that MoEPlan, with only two plans generated, takes less inference time, while still producing more efficient plans than other learned plan-selection optimizers.

Memory disaggregation has emerged as a promising architecture for improving resource efficiency by decoupling the computing and memory resources. But building efficient range indices in such an architecture faces three critical challenges: (1) coarse-grained concurrency control schemes for coordinating concurrent read/write operations with node splitting incur high contention under the skewed and write-intensive workloads; (2) existing data layouts fail to balance consistency verification and hardware acceleration via SIMD (Single Instruction Multiple Data); and (3) naive caching schemes struggle to adapt to rapidly changing access patterns. To address these challenges, we proposeMariana, a memory-disaggregated skiplist index that integrates three key innovations. First, it uses a fine-grained (i.e., entry-level) latch mechanism combined with dynamic node resizing to minimize the contention and splitting frequency. Second, it employs a tailored data layout for leaf node, which separates keys and values to enable SIMD acceleration while maintaining consistency checks with minimal write overhead. Third, it implements an adaptive caching strategy that tracks node popularity in real-time to optimize network bandwidth utilization during the index traversal. Experimental results show thatMarianaachieves$1.7\times$higher throughput under write-intensive workloads and reduces the P90 latency by 23% under the read-intensive workloads, when comparing to the state-of-the-art indices on disaggregated memory.

In distributed storage systems, replication and erasure code (EC) are common methods for data redundancy. Compared with replication, EC has better storage efficiency, but suffers higher overhead in update. Moreover, consistency and reliability problems caused by concurrent updates bring new challenges to applications of EC. Many works focus on optimizing the EC solution, including algorithm optimization, novel data update method, and so on, but lack the solutions for consistency and reliability problems. In this paper, we introduce a storage system that decouples data updating and EC encoding, namely, decoupled data updating and coding (DDUC), and propose a data placement policy that combines replication and parity blocks. For the ( N, M ) EC system, the data are placed as N groups of M +1 replicas, and redundant data blocks of the same stripe are placed in the parity nodes, so that the parity nodes can autonomously perform local EC encoding. Based on the above policy, a two-phase data update method is implemented in which data are updated in replica mode in phase 1, and the EC encoding is done independently by parity nodes in phase 2. This solves the problem of data reliability degradation caused by concurrent updates while ensuring high concurrency performance. It also uses persistent memory (PMem) hardware features of the byte addressing and eight-byte atomic write to implement a lightweight logging mechanism that improves performance while ensuring data consistency. Experimental results show that the concurrent access performance of the proposed storage system is 1.70–3.73 times that of the state-of-the-art storage system Ceph, and the latency is only 3.4%–5.9% that of Ceph.

NVMe (Non-Volatile Memory Express) is a protocol designed specifically for SSD (Solid State Drive), which has significantly improved the performance of SSD storage devices. However, the traditional kernel-space IO path hinders the performance of NVMe SSD devices. In this paper, a user-space raw file system (URFS) based on NVMe SSD is proposed. Through the design of the user-space multi-process shared cache, multiple applications can share access to SSD to reduce the amount of SSD access; NVMe-oriented log-free data layout and Multi-granularity IO queue elastic separation technology are used to improve system performance and throughput. Experiments show that, compared to traditional file systems, URFS performance is improved by more than 23% in CDN (Content Delivery Network) scenarios, and URFS performance is improved more in small file scenarios and read-intensive scenarios.