arXiv subjects

Fengyuan Ren

Publications and source records attributed to Fengyuan Ren.

10 recordsLinked to original sources

arXiv2026

Xkernel: Principled Performance Tunability of Operating System Kernels

The Linux kernel is permeated with constant values that are critical to system performance. Many of these constants, referred to as perf-consts, are magic numbers with brittle assumptions on hardware and workloads. Unfortunately, there is no capability of in-situ tuning of perf-const values on deployed kernels. This paper rethinks OS performance tunability. We present Xkernel, a system that offers a safe, efficient, and programmable interface for in-situ tuning of any perf-consts directly on a running kernel. Xkernel transforms any perf-const into a tunable knob on demand using a novel approach called Scoped Indirect Execution (SIE). SIE captures precise binary boundaries where a perf-const enters system state and redirects control to synthesized instructions that update the state as if new values were used. Xkernel goes beyond version atomicity when updating perf-consts to guarantee side-effect safety, a property notably absent in existing kernel update mechanisms. Case studies on various OS subsystems demonstrate significant performance benefits of tuning perf-consts which is made possible by Xkernel.

cs.OS↗

arXiv2026

Communication-Semantic-Aware RDMA Loss Recovery for QP-scalable Hyperscale AI Training

Current artificial intelligence (AI) infrastructures widely adopt Remote Direct Memory Access (RDMA) to support high-performance communication. Training trillion-parameter models involves frequent collective communication operations, such as All-Reduce and All-to-All, which generate intensive RDMA traffic. Existing RDMA deployments predominantly use the reliable connection (RC) model, where each process pair requires a dedicated queue pair (QP). This leads to poor scalability: since the RDMA-capable network interface card (RNIC) can cache only a few thousand QPs, excess entries trigger PCIe round-trip penalties. Meanwhile, global synchronization makes training sensitive to tail latency, where a few packet losses can delay iteration completion. To address these challenges, we propose Communication-Semantic-Aware Unreliable Datagram (CSA-UD), a novel RDMA loss recovery mechanism that combines scalability and reliability. CSA-UD decouples data transmission from loss recovery and dynamically adjusts the loss detection interval, accelerating tail recovery and exploiting the synchronization semantics of distributed training. It further supports multipath transmission and bitmap-guided reassembly, enabling high throughput without requiring lossless fabrics. Testbed experiments and ns-3 simulations show that CSA-UD significantly reduces tail latency under large-scale collective communication. Under high network load, it achieves better scalability than RC and over 30% lower 99th percentile flow completion times compared with counterparts.

cs.NI↗

arXiv2026

Q-StaR: A Quasi-Static Routing Scheme for NoCs

In networks-on-chip, static routing schemes are favored for their simplicity and predictability, but they cannot effectively balance network load due to the unawareness of runtime load distribution. Q-StaR discovers two factors (topology and traffic distribution) that determine the long-term trend of load distribution, and proposes N-Rank to extract this trend. The obtained information is used to guide BiDOR's route selection at runtime, thereby improving load balancing while retaining simplicity and predictability. Simulation validates that Q-StaR significantly outperforms the typical dimension-order routing (throughput under uniform traffic improved by 42.9\%, and mean/maximum latency under realistic workloads reduced by 86.4\%/95.3\%).

cs.NI↗

arXiv2025

An Extensible Software Transport Layer for GPU Networking

Fast-evolving machine learning (ML) workloads have increasing requirements for networking. However, host network transport on RDMA NICs is hard to evolve, causing problems for ML workloads. For example, single-path RDMA traffic is prone to flow collisions that severely degrade collective communication performance. We present UCCL, an extensible software transport layer to evolve GPU networking. UCCL decouples the data path and control path of existing RDMA NICs and efficiently runs the control-path transport on host CPUs. This software extensibility brings in transport innovations that cannot be achieved in hardware for ML workloads, e.g., a multipath transport to resolve flow collisions. ML collectives atop UCCL achieve up to 4.5x higher performance compared to existing RDMA NICs.

cs.NI↗

arXiv2025

Occamy: A Preemptive Buffer Management for On-chip Shared-memory Switches

Today's high-speed switches employ an on-chip shared packet buffer. The buffer is becoming increasingly insufficient as it cannot scale with the growing switching capacity. Nonetheless, the buffer needs to face highly intense bursts and meet stringent performance requirements for datacenter applications. This imposes rigorous demand on the Buffer Management (BM) scheme, which dynamically allocates the buffer across queues. However, the de facto BM scheme, designed over two decades ago, is ill-suited to meet the requirements of today's network. In this paper, we argue that shallow-buffer switches, intense bursts, along with dynamic traffic call for a highly agile BM that can quickly adjust the buffer allocation as traffic changes. However, the agility of the current BM is fundamentally limited by its non-preemptive nature. Nonetheless, we find that preemptive BM, considered unrealizable in history, is now feasible on modern switch chips. We propose Occamy, a preemptive BM that can quickly adjust buffer allocation. Occamy utilizes the redundant memory bandwidth to actively reclaim and reallocate the over-allocated buffer. Testbed experiments and large-scale simulations show that Occamy can improve the end-to-end performance by up to ~55%.

cs.NI↗

arXiv2022

Energy-Sensitive Trajectory Design and Restoration Areas Allocation for UAV-Enabled Grassland Restoration

Grassland restoration is a critical means to safeguard grassland ecological degradation. To alleviate the extensive human labors and boost the restoration efficiency, UAV is promising for its fully automatic capability yet still waits to be exploited. This paper progresses this emerging technology by explicitly considering the realistic constraints of the UAV and the grassland degradation while planning the grassland restoration. To this end, the UAV-enabled restoration process is first mathematically modeled as the maximization of restoration areas of the UAV under the limited battery energy of UAV, the grass seeds weight, the number of restored areas, and the corresponding sizes. Then we analyze that, by considering these constraints, this original problem emerges two conflict objectives, namely the shortest flight path and the optimal areas allocation. As a result, the maximization of restoration areas turns out to be a composite of a trajectory design problem and an areas allocation problem that are highly coupled. From the perspective of optimization, this requires solving two NP-hard problems of both the traveling salesman problem (TSP) and the multidimensional knapsack problem (MKP) at the same time. To tackle this complex problem, we propose a cooperative optimization algorithm, called CHAPBILM, to solve those two problems interlacedly by utilizing the interdependencies between them. Multiple simulations verify the conflicts between the trajectory design and areas allocation. The effectiveness of the cooperative optimization algorithm is also supported by the comparisons with traditional optimization methods which do not utilize the interdependencies between the two problems. As a result, the proposed algorithm successfully solves the multiple simulation instances in a near-optimal way.

cs.NE↗

arXiv2018

Micro Congestion Control: Every Flow Deserves a Second Chance

Today, considerable Internet traffic is sent from the datacenter and heads for users. The characteristics of connections served by servers in datacenters are usually diverse and varied over time, with continuous upgrades in network infrastructure and user devices. As a result, a specific congestion control algorithm hardly accommodates the heterogeneity and performs well in various scenarios. In this work, we present Micro Congestion Control (MCC) --- a novel framework for Internet congestion control. With MCC, diverse algorithms can be assigned purposely to connections in one server to adapt to heterogeneity, and different algorithms can be chosen in each connection's life cycle to keep pace with the dynamic of network. We design and implement MCC in Linux, and the experiments validate that MCC is capable of smoothly switching among various candidate algorithms on the fly to achieve potential performance gain in the real world. Meanwhile, the overheads introduced by MCC are moderate and acceptable.

cs.NI↗

arXiv2016

Isolating Mice and Elephant in Data Centers

Data centers traffic is composed by numerous latency-sensitive "mice" flows, which is consisted of only several packets, and a few throughput-sensitive "elephant" flows, which occupy more than 80% of overall load. Generally, the short-lived "mice" flows induce transient congestion and the long-lived "elephant" flows cause persistent congestion. The network congestion is a major performance inhibitor. Conventionally, the hop-by-hop and end-to-end flow control mechanisms are employed to relief transient and persistent congestion, respectively. However, in face of the mixture of elephants and mice, we find the hybrid congestion control scheme including hop-by-hop and end-to-end flow control mechanisms suffers from serious performance impairments. As a step further, our in-depth analysis reveals that the hybrid scheme performs poor at latency of mice and throughput of elephant. Motivated by this understanding, we argue for isolating mice and elephants in different queues, such that the hop-by-hop and end-to-end flow control mechanisms are independently imposed to short-lived and long-lived flows, respectively. Our solution is readily-deployable and compatible with current commodity network devices and can leverage various congestion control mechanisms. Extensive simulations show that our proposal of isolation can simultaneously improve the latency of mice by at least 30% and the link utilization to almost 100%.

cs.NI↗

arXiv2016

Micro-burst in Data Centers: Observations, Implications, and Applications

Micro-burst traffic is not uncommon in data centers. It can cause packet dropping, which results in serious performance degradation (e.g., Incast problem). However, current solutions that attempt to suppress micro-burst traffic are extrinsic and ad hoc, since they lack the comprehensive and essential understanding of micro-burst's root cause and dynamic behavior. On the other hand, traditional studies focus on traffic burstiness in a single flow, while in data centers micro-burst traffic could occur with highly fan-in communication pattern, and its dynamic behavior is still unclear. To this end, in this paper we re-examine the micro-burst traffic in typical data center scenarios. We find that evolution of micro-burst is determined by both TCP's self-clocking mechanism and bottleneck link. Besides, dynamic behaviors of micro-burst under various scenarios can all be described by the slope of queue length increasing. Our observations also implicate that conventional solutions like absorbing and pacing are ineffective to mitigate micro-burst traffic. Instead, senders need to slow down as soon as possible. Inspired by the findings and insights from experimental observations, we propose S-ECN policy, which is an ECN marking policy leveraging the slope of queue length increasing. Transport protocols utilizing S-ECN policy can suppress the sharp queue length increment by over 50%, and reduce the 99th percentile of query completion time by ~20%.

cs.NI↗

arXiv2014

Scale Congestion Control to Ultra-High Speed Ethernet

Currently, Ethernet is broadly used in LAN, datacenter and enterprise networks, storage networks, high performance computing networks and so on. Along with the popularity of Ethernet comes the requirement of enhancing Ethernet with congestion control. On the other hand, Ethernet speed extends to 40Gbps and 100Gbps recently, and even 400Gbps in the near future. The ultra-high speed requires congestion control algorithms to adapt to the broad changes of bandwidth, and highlights the impacts of small delay by enlarging the bandwidth delay product. The state-of-art standard QCN is heuristically designed for the 1Gbps and 10Gbps Ethernet, and unaware of the challenges accompanying the ultra-high speed. To scale congestion control to ultra-high speed Ethernet, we propose the Adaptive Sliding Mode (ASM) congestion control algorithm, which is simple, stable, has fast and smooth convergence process, can tolerate the impacts of delay and adapt to the wide changes of bandwidth. Real experiments and simulations confirm these good properties and show that ASM outperforms QCN. Designing ASM, we find that the derivative of queue length is helpful to rate adjustment because it reflects the difference between bandwidth and aggregated sending rate. We also argue for enforcing congestion control system staying at the congestion boundary line, along which it automatically slides to stable point. These insights are also valuable to develop other congestion control algorithms in ultra-high speed networks.

cs.NI↗