Research Team Led by Professor Park Sung-yong of the Department of Computer Science and Engineering, Wins Paper Acceptance at IEEE ICDE 2026
Byun Hong-soo (co-first author, Ph.D.), Lee Seung-jae (co-first author, Master’s student), Yoo Hong-hyun (Integrated Master’s-Ph.D. student), Kim Myung-jun (Master’s student), and Professor Park Sung-yong (corresponding author) from the Data Intensive AI Computing and Systems Laboratory (DISCOS) have had their paper titled “Resystance: Unleashing Hidden Performance of Compaction in LSM-trees via eBPF” accepted for publication at the IEEE International Conference on Data Engineering (ICDE) 2026.
Storage technology has advanced dramatically over the past few years. The latest NVMe SSDs provide multi-GB/s bandwidth and microsecond-level latency, overcoming the performance limitations of legacy disk hardware. Paradoxically, however, as hardware becomes faster, software inefficiencies become more apparent. The bottleneck in storage systems is now shifting from hardware to the software stack, comprising the operating system, file system, and system call interfaces.
These structural bottlenecks are particularly severe in LSM-tree-based NoSQL databases. RocksDB, a representative LSM-tree-based key-value store (LSM-KVS), repeatedly performs compaction in the background to maintain high write throughput. In this study, we conducted a system-call-level analysis and determined that compaction dominates the internal system calls within the LSM-tree. In particular, the study identified that the pread() system call, which occurs during the reading of numerous data blocks, is the primary source of the software stack bottleneck. These results demonstrate that, in modern high-speed storage environments, the limitations of LSM-trees reside not in the hardware, but at the system software boundary between the operating system and the database.
Motivated by these software stack bottlenecks, this study proposes Resystance, which mitigates overhead rather than requiring new hardware or modifying the LSM-tree structure. Resystance leverages Linux’s eBPF and io_uring to offload the core I/O and merge paths of compaction to the kernel, thereby drastically reducing unnecessary user-to-kernel boundary crossings and system calls. This approach eliminates bottlenecks by controlling the execution path within the OS kernel while preserving the existing database engine’s architecture.
Key experimental results demonstrate that Resystance achieved a near-total 99% reduction in system calls during compaction and slashed compaction time by approximately 50%. Under write-intensive workloads, it delivered up to a 75% increase in throughput and a 40% reduction in p99 latency. This feat is especially noteworthy as these gains were realized solely through pure system-level software optimizations, without relying on additional hardware acceleration.
Foundational technologies such as the OS I/O path, kernel programmable frameworks, and system call structures play a pivotal role in the performance of applications, including databases. This study demonstrates that as storage devices become faster, the design of the system software bridging the OS and the database becomes increasingly critical. The competitive edge of data systems now depends less on simply adopting faster hardware and more on how precisely software bottlenecks are analyzed and how the boundary between the OS and the database is architected. At this turning point, Resystance represents a significant research milestone that sets the course for next-generation data infrastructure.
Lee Seung-jae, a master’s student and co-first author of the paper, shared his thoughts: “As kernel-programmable frameworks continue to be a vibrant area of research, I believe we have opened the door to their application across diverse fields. I am deeply grateful to my colleagues for their hard work and to my advisor for the guidance that made this achievement possible.”
The IEEE International Conference on Data Engineering (ICDE) is a prestigious global forum in the field of data engineering, dedicated to sharing cutting-edge research in database systems, large-scale data management, and data-intensive computing. ICDE is recognized as a top-tier international conference with a BK21 Computer Science adjusted impact factor (IF) of 3 and is ranked as a premier conference in the software field by the Korean Institute of Information Scientists and Engineers (KIISE). This year’s conference is scheduled to be held in Montreal, Canada, from May 4th to 9th.
Title: Resystance: Unleashing Hidden Performance of Compaction in LSM-trees via eBPF
Authors: Byun Hong-su (Co-first author), Lee Seung-jae (Co-first author), Yoo Hong-hyun (Third author), Kim Myeong-jun (Fourth author), Park Sung-yong (Corresponding author)
[SEO 키워드]
IEEE ICDE 2026, LSM-tree Compaction, eBPF Optimization
[Summary]
The research team secured a paper acceptance at IEEE ICDE 2026 for developing Resystance, a framework that leverages eBPF Optimization to eliminate software bottlenecks in LSM-tree Compaction.
Storage technology has advanced dramatically over the past few years. The latest NVMe SSDs provide multi-GB/s bandwidth and microsecond-level latency, overcoming the performance limitations of legacy disk hardware. Paradoxically, however, as hardware becomes faster, software inefficiencies become more apparent. The bottleneck in storage systems is now shifting from hardware to the software stack, comprising the operating system, file system, and system call interfaces.
These structural bottlenecks are particularly severe in LSM-tree-based NoSQL databases. RocksDB, a representative LSM-tree-based key-value store (LSM-KVS), repeatedly performs compaction in the background to maintain high write throughput. In this study, we conducted a system-call-level analysis and determined that compaction dominates the internal system calls within the LSM-tree. In particular, the study identified that the pread() system call, which occurs during the reading of numerous data blocks, is the primary source of the software stack bottleneck. These results demonstrate that, in modern high-speed storage environments, the limitations of LSM-trees reside not in the hardware, but at the system software boundary between the operating system and the database.
Motivated by these software stack bottlenecks, this study proposes Resystance, which mitigates overhead rather than requiring new hardware or modifying the LSM-tree structure. Resystance leverages Linux’s eBPF and io_uring to offload the core I/O and merge paths of compaction to the kernel, thereby drastically reducing unnecessary user-to-kernel boundary crossings and system calls. This approach eliminates bottlenecks by controlling the execution path within the OS kernel while preserving the existing database engine’s architecture.
Key experimental results demonstrate that Resystance achieved a near-total 99% reduction in system calls during compaction and slashed compaction time by approximately 50%. Under write-intensive workloads, it delivered up to a 75% increase in throughput and a 40% reduction in p99 latency. This feat is especially noteworthy as these gains were realized solely through pure system-level software optimizations, without relying on additional hardware acceleration.
Foundational technologies such as the OS I/O path, kernel programmable frameworks, and system call structures play a pivotal role in the performance of applications, including databases. This study demonstrates that as storage devices become faster, the design of the system software bridging the OS and the database becomes increasingly critical. The competitive edge of data systems now depends less on simply adopting faster hardware and more on how precisely software bottlenecks are analyzed and how the boundary between the OS and the database is architected. At this turning point, Resystance represents a significant research milestone that sets the course for next-generation data infrastructure.
Lee Seung-jae, a master’s student and co-first author of the paper, shared his thoughts: “As kernel-programmable frameworks continue to be a vibrant area of research, I believe we have opened the door to their application across diverse fields. I am deeply grateful to my colleagues for their hard work and to my advisor for the guidance that made this achievement possible.”
The IEEE International Conference on Data Engineering (ICDE) is a prestigious global forum in the field of data engineering, dedicated to sharing cutting-edge research in database systems, large-scale data management, and data-intensive computing. ICDE is recognized as a top-tier international conference with a BK21 Computer Science adjusted impact factor (IF) of 3 and is ranked as a premier conference in the software field by the Korean Institute of Information Scientists and Engineers (KIISE). This year’s conference is scheduled to be held in Montreal, Canada, from May 4th to 9th.
Title: Resystance: Unleashing Hidden Performance of Compaction in LSM-trees via eBPF
Authors: Byun Hong-su (Co-first author), Lee Seung-jae (Co-first author), Yoo Hong-hyun (Third author), Kim Myeong-jun (Fourth author), Park Sung-yong (Corresponding author)
[SEO 키워드]
IEEE ICDE 2026, LSM-tree Compaction, eBPF Optimization
[Summary]
The research team secured a paper acceptance at IEEE ICDE 2026 for developing Resystance, a framework that leverages eBPF Optimization to eliminate software bottlenecks in LSM-tree Compaction.