arXiv subjects

Yijun Sun

Publications and source records attributed to Yijun Sun.

5 recordsLinked to original sources

arXiv2026

Multi-stage Flow Scheduling for LLM Serving

Meeting stringent Time-To-First-Token (TTFT) requirements is crucial for LLM applications. To improve efficiency, modern LLM serving systems adopt disaggregated architectures with diverse parallelisms, introducing complex multi-stage workflows involving reusable KV-block retrieval, collective communication, and P2D transfer. Flows from dependent stages overlap within and across requests on shared bottleneck links, making TTFT highly susceptible to network contention and necessitating stage-aware scheduling. Unfortunately, most existing works schedule flows in a stage-agnostic manner, leading to uncoordinated contention that constitutes a primary cause of SLO violations. In this paper, we present MFS, a holistic multi-stage flow scheduling mechanism designed to maximize TTFT SLO attainment. At its core, MFS approximates the Least-Laxity-First (LLF) scheduling policy without requiring precise knowledge of a request's remaining slack. It achieves this through a Defer-and-Promote principle implemented through a Reverse Multi-Level Queue (RMLQ) structure. By dynamically promoting task precedence as effective laxity diminishes, MFS prioritizes flows with less laxity while preventing requests with loose SLOs from prematurely consuming network bandwidth. We implement MFS as a pluggable module integrated into vLLM, and evaluate it on a 8-server, 32-GPU testbed as well as through large-scale simulations. Our results demonstrate that MFS effectively outperforms state-of-the-art baselines, improving the TTFT SLO attainment by 1.2x--2.4x.

cs.NI↗

arXiv2025

FLASH-FHE: A Heterogeneous Architecture for Fully Homomorphic Encryption Acceleration

While many hardware accelerators have recently been proposed to address the inefficiency problem of fully homomorphic encryption (FHE) schemes, none of them is able to deliver optimal performance when facing real-world FHE workloads consisting of a mixture of shallow and deep computations, due primarily to their homogeneous design principle. This paper presents FLASH-FHE, the first FHE accelerator with a heterogeneous architecture for mixed workloads. At its heart, FLASH-FHE designs two types of computation clusters, ie, bootstrappable and swift, to optimize for deep and shallow workloads respectively in terms of cryptographic parameters and hardware pipelines. We organize one bootstrappable and two swift clusters into one cluster affiliation, and present a scheduling scheme that provides sufficient acceleration for deep FHE workloads by utilizing all the affiliations, while improving parallelism for shallow FHE workloads by assigning one shallow workload per affiliation and dynamically decomposing the bootstrappable cluster into multiple swift pipelines to accelerate the assigned workload. We further show that these two types of clusters can share valuable on-chip memory, improving performance without significant resource consumption. We implement FLASH-FHE with RTL and synthesize it using both 7nm and 14/12nm technology nodes, and our experiment results demonstrate that FLASH-FHE achieves an average performance improvement of $1.4\times$ and $11.2\times$ compared to state-of-the-art FHE accelerators CraterLake and F1 for deep workloads, while delivering up to $8.0\times$ speedup for shallow workloads due to its heterogeneous architecture.

cs.AR↗

arXiv2025

MixNet: A Runtime Reconfigurable Optical-Electrical Fabric for Distributed Mixture-of-Experts Training

Mixture-of-Expert (MoE) models outperform conventional models by selectively activating different subnets, named experts, on a per-token basis. This gated computation generates dynamic communications that cannot be determined beforehand, challenging the existing GPU interconnects that remain static during the distributed training process. In this paper, we advocate for a first-of-its-kind system, called MixNet, that unlocks topology reconfiguration during distributed MoE training. Towards this vision, we first perform a production measurement study and show that the MoE dynamic communication pattern has strong locality, alleviating the requirement of global reconfiguration. Based on this, we design and implement a regionally reconfigurable high-bandwidth domain on top of existing electrical interconnects using optical circuit switching (OCS), achieving scalability while maintaining rapid adaptability. We have built a fully functional MixNet prototype with commodity hardware and a customized collective communication runtime that trains state-of-the-art MoE models with in-training topology reconfiguration across 32 A100 GPUs. Large-scale packet-level simulations show that MixNet delivers comparable performance as the non-blocking fat-tree fabric while boosting the training cost efficiency (e.g., performance per dollar) of four representative MoE models by 1.2x-1.5x and 1.9x-2.3x at 100 Gbps and 400 Gbps link bandwidths, respectively.

cs.NI↗

arXiv2024

Metasurface spectrometers beyond resolution-sensitivity constraints

Optical spectroscopy plays an essential role across scientific research and industry for non-contact materials analysis1-3, increasingly through in-situ or portable platforms4-6. However, when considering low-light-level applications, conventional spectrometer designs necessitate a compromise between their resolution and sensitivity7,8, especially as device and detector dimensions are scaled down. Here, we report on a miniaturizable spectrometer platform where light throughput onto the detector is instead enhanced as the resolution is increased. This planar, CMOS-compatible platform is based around metasurface encoders designed to exhibit photonic bound states in the continuum9, where operational range can be altered or extended simply through adjusting geometric parameters. This system can enhance photon collection efficiency by up to two orders of magnitude versus conventional designs; we demonstrate this sensitivity advantage through ultra-low-intensity fluorescent and astrophotonic spectroscopy. This work represents a step forward for the practical utility of spectrometers, affording a route to integrated, chip-based devices that maintain high resolution and SNR without requiring prohibitively long integration times.

physics.optics↗

arXiv2016

A Novel Regularized Principal Graph Learning Framework on Explicit Graph Representation

Many scientific datasets are of high dimension, and the analysis usually requires visual manipulation by retaining the most important structures of data. Principal curve is a widely used approach for this purpose. However, many existing methods work only for data with structures that are not self-intersected, which is quite restrictive for real applications. A few methods can overcome the above problem, but they either require complicated human-made rules for a specific task with lack of convergence guarantee and adaption flexibility to different tasks, or cannot obtain explicit structures of data. To address these issues, we develop a new regularized principal graph learning framework that captures the local information of the underlying graph structure based on reversed graph embedding. As showcases, models that can learn a spanning tree or a weighted undirected $\ell_1$ graph are proposed, and a new learning algorithm is developed that learns a set of principal points and a graph structure from data, simultaneously. The new algorithm is simple with guaranteed convergence. We then extend the proposed framework to deal with large-scale data. Experimental results on various synthetic and six real world datasets show that the proposed method compares favorably with baselines and can uncover the underlying structure correctly.

cs.AI↗