Jinlei Hu

Modern vector search systems increasingly depend on highcapacity and cost-efficient NVMe SSDs for storage, as data dimensionality and scale continue to grow. Graph-based indices are widely adopted for their superior search accuracy; however, the graph traversal triggers intensive random I/O on SSDs, significantly degrading search performance. Existing graph-based retrieval methods attempt to reduce I/O overhead through page expansion. However, these approaches introduce two key limitations: 1) Insufficient page expansion. As vector dimensions grow, fewer nodes fit within a single SSD page, reducing the number of accessible nodes for expansion and thus limiting traversal efficiency. 2) Inefficient utilization of I/O time. Existing candidate pool expansion strategy exhibits diminishing returns in reducing I/O overhead, while its computation increases linearly, resulting in inefficient utilization of I/O time. To address these issues, we propose Giant, an I/O-optimized graph index for high-dimensional vector search based on a novel page group expansion mechanism. This mechanism is supported by two core techniques: a memory-resident group index enabling deeper node expansion, and a page expansion list offering diverse supplementary candidates. Experiments demonstrate that Giant reduces average search latency by 12%–43% and achieves 1.10×–1.41× higher throughput compared with state-of-the-art methods.

The host-side and append-only zone interface offers new opportunities for existing key-value stores (KVSs), especially in reducing the flash-layer write amplification. While existing works focused on leveraging the zone interface for LSM-Tree-based KVSs, the potential advantages for B+-Tree-based KVSs remain under-explored. Through in-depth experimental observations, we identify three key opportunities for B+-Tree-based KVSs: superior read performance, flash-friendly append operations, and the ability to directly manage flash media.However, existing B+-Tree-based KVSs designed for block SSDs rely on inefficient file system layers to adapt to the idiosyncratic zone interface, and achieve reduced write amplification by seriously sacrificing read performance. We proposed a write-optimized B+-Tree-based key-value store ZKV, designed to minimize write amplification through the zone interface without sacrificing read performance. ZKV employs a three-level buffer structure to enable its core index structure Z+-Tree to be zone-interface efficient, including a tree-level merged filter, a leaf-level adaptive delta buffer, and an efficient zone-level management module ZFlusher.Through the three-level buffer structure, Z+-Tree effectively leverages the zone interface by converting smallsize random write operations into flash-friendly large-block append operations. Our evaluations on a real ZNS SSD device demonstrate that ZKV achieves up to 3.12x/2.53x higher insert/read throughput than the current buffered-B+-Tree-based KVSs under YCSB workloads. Under three additional real-world workloads, ZKV matches the write performance of LSM-Tree-based RocksDB while delivering the read performance of B+-Tree-based WiredTiger.

Persistent memory (PM) offers a compelling combination of durability and near-DRAM performance, but it also introduces new challenges for hashing indexes. Existing persistent hashing designs prioritize resizing efficiency at the expense of increased query latency, losing the key advantage of hash tables. This paper introduces R2Hash, a persistent hash index redesigned from the persistent cache-line hash table, to balance high read performance with efficient resizing. R2Hash is guided by a migration rule, enabling it to meet both goals through two main contributions: (i) a cooperative and lowoverhead resizing strategy based on split-order hashing, and (ii) a shift-aware search combined with a two-layer bucket layout that enables lock-free reads with only one PM access on average. Furthermore, R2Hash provides log-free consistency and a nonblocking recovery mechanism. Experimental results demonstrate that R2Hash achieves up to 8.1× higher search throughput and 7.5× higher insert throughput compared to other persistent hash indexes across a range of workloads

Persistent Memory (PM) offers both byte-address ability and non-volatility, making it well-suited for accelerating B+-Tree indexes. However, existing persistent B+-Tree indexes face significant performance challenges due to high structural modification operation (SMO) overhead. SMOs often result in costly item migrations and increased tail latency, which severely degrade the overall performance. In this paper, we present SSTree, a high-performance B+– Tree index specifically optimized to address SMO overhead. SSTree introduces three key innovations: (i) efficient leaf node expansion using a list of subnodes to postpone expensive node splits, (ii) delegated fingerprints to speed up search operations across subnodes, and (iii) proactive subnode compaction that employs out-of-place updates to optimize item organization. Our evaluation demonstrates that SSTree delivers up to 4.38× higher write throughput and up to 62× lower tail latency compared to state-of-the-art persistent B+-Tree indexes.

Tree index structures are widely employed in modern storage systems to support high-performance queries. Persistent memory (PM) brings a new opportunity and challenge for tree indexes. Among persistent tree indexes, we find the radix tree is more suitable than B-Tree for the byte-ability of PM. However, the hierarchy of radix remains excessively high, resulting in high read latency. Node splitting imposes a significant overhead on PM. To address these challenges, we propose RWORT, a read and write optimized radix tree for PM. The key focus of RWORT is to minimize random access in PM and provide efficient write operations. RWORT proposed a hierarchical compression mechanism to significantly reduce the tree height. Additionally, RWORT incorporates mini bloom filters to reduce unnecessary access on PM. For efficient write operations, RWORT uses the lazy split flag and the double-linked pointers to reduce the critical path delay. Furthermore, RWORT introduces a low-overhead ring-based bit tree allocator that improves allocation efficiency on PM. Our experiments show that RWORT improves up to 1.62x/4.91x respectively compared to the state-of-the-art radix tree/B-Tree. RWORT also exhibits higher performance in real-world storage systems such as Memcached.