FAST '27 Spring Accepted Papers | USENIX

FAST '27 has two submission deadlines. Prepublication versions of the spring accepted papers and their abstracts are available below.
Jaehoon Shim, Hyungjoon Kwon, Hyungsoo Jung, and Jin-Soo Kim, Seoul National University
Computational storage devices (CSDs) promise to reduce host--storage data movement by executing computation inside the storage device. This is especially appealing for LSM-tree-based key--value stores, whose reads and background maintenance move large amounts of data through the host. In practice, however, in-storage LSM-tree operations often fail to deliver their full benefit. Existing CSDs largely follow a host-managed, copy-then-execute model that drives in-storage operators through synchronous or weakly overlapped I/O paths. This serializes compute-flash interactions, underutilizes device-side compute resources, and forces host storage software to handle complex coordination logic. As a result, the expected benefits of in-storage processing are often limited or even negated on realistic workloads.
We argue that the missing abstraction is asynchronous execution with device-resident dependency management. We present FLAX, a computational storage architecture that makes asynchronous, overlapped in-device I/O execution a first-class design goal for LSM-tree operations. FLAX tracks fine-grained dependencies inside the device, allowing the host to issue large batches of work while the CSD schedules and executes commands asynchronously. To further improve utilization, FLAX combines adaptive data loading with dynamic on-device buffer management, enabling higher concurrency under limited device resources. We integrate FLAX into RocksDB and offload both compaction and reads. FLAX improves throughput by up to 3.1× over host-only execution and 1.8× over a host-managed CSD baseline.
Raúl Gracia-Tinedo and Hossam Elghamry, Dell Technologies
Modern event streaming systems (e.g., Apache Kafka, Apache Pulsar) provide support for storage tiering: the ability to move cold stream data to an external storage service (e.g., AWS S3). In this paper, we identify that there is an opportunity for exploiting advanced data management on data streams during the storage tiering process. To this end, we present Nexus: a data management mesh that transparently mediates storage operations between streaming systems and external storage. The main abstraction in Nexus is the streamlet: a data management function —e.g., processing, routing, caching— executed inline on a chunk of tiered stream data. Administrators can deploy streamlets across heterogeneous infrastructures and orchestrate them via policies. We demonstrate Nexus via stateless and stateful streamlets. We show that Nexus achieves up to 3.9× higher data reduction than built-in compression in streaming systems, supports privacy-aware data routing (reducing metadata reads by 98%), and allows for complex health-related data management pipelines. All this is achieved transparently across multiple streaming systems (Kafka, Pulsar, Pravega) and incurring low interception overhead (< 5%).
Xiuqi Huang, Zhejiang University; Qingsong Zhu, Yaohua Tang, Honghao Wang, Xin Wei, Jinxin Liu, Guanyu Liu, Zongjia Chen, Tianyue Lan, Yunlong Wang, Hang Zhang, Rusong Jiao, Yuanwang Yue, Jie Ding, Xiaodong Luo, Zijin Han, Huan Sun, and Maoyin Liu, ByteDance Inc.; Wei Chen, Zhejiang University; Chengyu Dong and Rui Shi, ByteDance Inc.
Heterogeneous cloud storage relies on tiering to balance performance and cost. However, intra- and inter-node tiering face three major challenges: limited flexibility in redundancy and scaling; excessive disk and network traffic amplification; inefficiency in handling small I/Os. We propose ByteTier, a pool-based tiering system. Based on decoupled performance and capacity pools, ByteTier designs a cross-pool two-phase write mechanism: writes are cached and stored in the performance pool in the foreground, and re-encoded from the cache to the capacity pool in the background, providing flexibility and reducing disk and network overhead. Within this mechanism, ByteTier performs a workload-aware I/O aggregation method, adaptively merging small writes and batching them into device-friendly patterns to enhance throughput and reduce latency. ByteTier has been deployed in production at ByteDance for three years, serving diverse workloads. Evaluation shows that it significantly reduces both average and P99 write latency while improving write and read throughput.
Xiaojun Guo, Huazhong University of Science an Technology; Guangjie Xing, Hua Wang, and Ke Zhou, Huazhong University of Science and Technology (HUST); Min Fu, Ming Xie, Fenqiang Yang, and Guangchao Yang, Tencent Technology (Shenzhen) Co., Ltd.; Jingyu Tang and Jinhui Xu, Huazhong University of Science and Technology (HUST)
Separating data by invalidation time is critical for mitigating garbage collection (GC) overhead in log-structured storage. In this paper, we establish a formal model and derive an OP-Minimized Oracle that optimizes separation thresholds for a theoretical Write Amplification Factor (WAF) of 1 under limited physical streams. Guided by this Oracle, we then design MINOS (MINimized Over-provisioning Separation), a practical online approximation that classifies writes into fine-grained virtual streams, tracks their invalidation behaviors, and adaptively clusters them into limited physical streams. Across 170 workloads totaling 500 TiB of writes, MINOS reduces GC writes by 15.0%--40.5% and improves throughput by up to 14.2% over state-of-the-art approaches, while incurring negligible CPU and memory overhead.
Yongxin Zhao, Yuxin Sun, and Shenglin Zhang, Nankai University; Xiang Liu, Huawei Cloud; Jiaqi Luan, Runzhou Wang, Wenwei Gu, Yongqian Sun, and Yidan Li, Nankai University; Dan Pei, Tsinghua University
Fail-slow failures, characterized by prolonged performance degradation without complete system crashes, pose a significant threat to the reliability of large-scale cloud