Breaking Barriers in Atomic Scaling: A Hardware-Software-Collaborated Framework to Deconstruct RDMA Atomic
Guangyang Deng, Qiangsheng Su, Zhirong Shen, Qing Wang, Yina Lv, Ronglong Wu, Jiwu Shu
Abstract
Remote Direct Memory Access (RDMA) Atomics are widely adopted to ensure correctness in distributed synchronization. However, their scalability is still seriously constrained by internal locking within RNICs. This paper provides systematic analysis that uncovers fundamental bottlenecks when scaling atomic operations. To break the constraints in RDMA Atomic scaling, we present Fusa, a framework that transparently coordinates server-side hardware (RNIC) and software (CPU) to accelerate atomic executions. Fusa integrates fine-grained contention detection with selective onloading, which executes uncontended operations in hardware while redirecting contended ones to software. Fusa further designs a consensus mechanism for strategy switching. We evaluate Fusa using microbenchmarks and unmodified RDMA-based systems, showing that Fusa improves throughput by up to 4.6×.
Ask Lune
Know where this paper stands in minutes, not an afternoon.
Lune can read the top-tier papers that cite it and the ones closest to it, and tells you what it claims, what came after and where it falls short, each point cited to the paper it comes from. No skimming, no AI slop.
Related papers
- Scalable RDMA-accelerated Distributed Locks with Shared Stream AbstractionMiao Cai, Junru Shen, Xiaojian Liao, Rong Gu et al.EuroSys 2026
- Fast and Scalable Selective Retransmission for RDMAPeihao Huang, Guo Chen, Xin Zhang, Can Liu et al.INFOCOM 2025 · 6 citations
- The nanoPU: A Nanosecond Network Stack for DatacentersStephen Ibanez, Alex Mallery, Serhat Arslan, Theo Jepsen et al.OSDI 2021 · 74 citations
- Tlaloc: A Generic Multipath Load Balancing for RoCEHuimin Luo, Jiao Zhang, Yongchen Pan, Tian Pan et al.INFOCOM 2026
- PeRF: Preemption-enabled RDMA FrameworkSugi Lee, Mingyu Choi, Ikjun Yeom, Younghoon KimUSENIX ATC 2024 · 3 citations
