FAST '26 Spring Accepted Papers

Approximate Nearest Neighbor Search (ANNS) is widely used in various scenarios. For billion-scale ANNS, on-disk graph-based indexes, which organize the vectors as a graph and store them on disk, are favored for their performance and cost-efficiency. However, existing indexes can not maintain a stable search performance while inserting new vectors.

In this paper, we propose to use direct insert, which directly inserts vectors into the on-disk index, rather than buffering them in memory and merging them to disk in batches like existing systems. This approach can even out the interference of insert with frontend search, thus stabilizing the performance. We evaluate direct insert by integrating it into a billion-scale graph-based ANNS index named OdinANN. With a fixed insert rate, OdinANN outperforms state-of-the-art ANNS indexes in search latency and throughput, and it consistently shows stable performance in billion-scale vector datasets.

With the development of high-performance RDMA network and modern storage devices, disaggregated NVMe SSDs have become increasingly popular due to the high resource utilization and superior performance. Although existing kernel file systems (e.g., Ext4) can directly access the disaggregated SSD via the NVMe-over-RDMA protocol, the data path suffers from heavy kernel stack overhead and thus degraded performance. The extra networking latency introduced in accessing remote storage also prevents file system from achieving scalable concurrent accesses and efficient failure-atomic IO. In this paper, we present CETOFS, a high-performance file system with host-server collaboration for disaggregated NVMe SSD. CETOFS designs a userspace-kernel collaborative architecture to place data plane entirely in userspace meanwhile separate permission checking from in-kernel control plane. Then CETOFS exploits the processing capability of remote storage server to offload three tasks: permission checking, concurrency control, and failure-atomic IO guaranteeing. The offloading mechanisms greatly reduce the networking overhead. We implement CETOFS and evaluate it against both kernel and userspace file systems. The evaluation shows CETOFS achieves high-performance data path that reduces latency by up to 52% for single-threaded file access and improves throughput by up to 19X for concurrent accesses.

Minimum Storage Regenerating (MSR) codes have strong potential for building efficient and reliable storage systems due to their excellent fault tolerance and low repair bandwidth. However, to meet MSR code constraints and optimize storage performance, systems often adopt large chunk sizes. This leads to significant I/O amplification during degraded reads, as entire chunks must be reconstructed to access a single object.

In this paper, we propose DRBoost, an approach that boosts degraded read performance in MSR-coded storage clusters by reducing repair bandwidth and eliminating access fragmentation for healthy data. DRBoost introduces three key techniques: (1) a partial-chunk reconstruction algorithm that reduces repair bandwidth by leveraging two forms of data reuse; (2) a reconstruction-friendly coding layout that improves reuse efficiency and accommodates objects of diverse sizes; and (3) a fragmentation-free storage layout that avoids unnecessary request splitting. Extensive experiments under various conditions and workloads show that DRBoost reduces degraded read latency by one to two orders of magnitude, significantly improving system responsiveness.

TapeOBS is an archive storage service offered by Huawei Cloud, which delivers high cost-efficiency by leveraging tape to store large volumes of archived data. Although tape boasts a low total cost of ownership, its inherent characteristics (e.g., a limited number of drives within a tape library) pose unique challenges when developing a large-scale distributed storage system. To address these challenges, we take a holistic approach in designing TapeOBS. At the high level, we introduce a fully asynchronous tape pool, which supports data scheduling and erasure coding in a batched manner, aligning with the features of tape hardware. Within a tape library, we design a tape-tailored local storage engine and incorporate techniques such as dedicated drives to optimize performance. TapeOBS began its gradual rollout at the end of 2022 and officially started serving customers in 2024. As of this writing, TapeOBS has stored hundreds of petabytes of raw user data.

Cloud-based performance monitoring timeseries systems are emerging due to their flexibility and pay-as-you-go capabilities. However, these systems encounter a major bottleneck in query performance, mainly attributed to the prolonged access latency of cloud storage and metadata redundancy of large number of timeseries. Thus, it is critical to optimize query performance within cloud environment and reduce metadata redundancy.

In this paper, we propose CloudTS, which is a novel timeseries data storage model with query optimization for cloud storage. CloudTS separately manages metadata and data, and introduces an efficient global metadata management for both space saving and query speedup. CloudTS also transparently supports the time-partitioned tag-based query model in performance monitoring timeseries systems. For metadata, a global tag dictionary is built to reduce metadata redundancy and a novel timeseries-tag mapping technique with a two-dimension bitmap is designed so the mapping of timeseries and tags can be efficiently accomplished to support tag-based queries. For data, the compressed data chunks are put into objects by timeseries group. We have implemented a fully functional prototype of CloudTS and evaluated it with production timeseries data and synthetic workloads based on Amazon S3. In comparison, Cortex, a cloud-based timeseries system widely adopted by industries, and Apache Parquet and JSON Time Series, two representative cloud storage formats, are utilized in the evaluation. Experimental results show that CloudTS can improve query performance by 1.37x on average compared with Cortex, and outperforms Apache Parquet and JSON Time Series as well.

Buffered I/O via page cache has been prevalently used by applications for decades due to its user-friendliness and high performance. However, the existing buffered I/O architecture fails to effectively utilize high-bandwidth Solid-State Drives (SSDs) caused by 1) costly page caching overused for buffering all incoming writes in the critical path, 2) the limited concurrency of page management, and 3) the high read-before-write penalty for partial-page writes.

This paper rearchitects buffered I/O and proposes a write-scrap buffering approach (WSBuffer) to remove the aforementioned shackles of buffered I/O on writes to proactively exploit fast SSDs while retaining all the advantages of buffered I/O on reads. WSBuffer first presents a novel memory-page buffering structure, scrap buffer, to efficiently buffer SSD-I/O unfriendly writes and expensive partial-page writes. WSBuffer further proposes a buffer-minimized data access mechanism to partially buffer small and unaligned parts of user writes via the scrap buffer while directly sending large and aligned parts to underlying SSDs. Finally, WSBuffer devises an opportunistic two-stage dirty-data flushing mechanism and a concurrent page management mechanism to achieve fluent and fast dirty-data flushing. The experimental results show that WSBuffer outperforms Linux file systems of EXT4, F2FS, BTRFS and XFS, as well as the state-of-the-art buffered I/O optimization of ScaleCache by up to 3.91X and 82.80X in throughput and latency respectively.

This paper presents Lockify, a novel distributed lock manager (DLM) for shared-disk file systems. Our key observation in shared-storage scenarios is that, for file or directory creation, lock acquisition overhead in the Linux kernel DLM increases with the number of clients, even in low-contention scenarios. Lockify minimizes this lock acquisition latency by avoiding unnecessary communication with remote directory nodes through self-owner notifications and asynchronous ownership management. We implement Lockify in the Linux kernel and evaluate its performance on real-world workloads using two representative shared-disk file systems, GFS2 and OCFS2. Our experimental results demonstrate that Lockify improves overall throughput by ~6.4× compared to the kernel DLM and O2CB, consistently across different numbers of clients.