arXiv subjects

Chunyang Zhang

Publications and source records attributed to Chunyang Zhang.

8 recordsLinked to original sources

arXiv2026

PSP: Low-Overhead Packet-Level Load Balancing for Stale-State and Bandwidth-Asymmetric Networks

With the rapid growth of large language model training and generative artificial intelligence services, data center networks face severe micro-burst traffic and high concurrency. Traditional hash-based flow-level load balancing cannot sense link states, leading to hash collisions, hotspot congestion, and tail latency in multipath Clos networks. Existing packet-level schemes are constrained by stale state information, high hardware complexity, and poor adaptation to heterogeneous links. To address these issues, this paper proposes probabilistic state-proportional (PSP) dispatching, a packet-level load balancing algorithm. Using a Band-based discrete state representation, PSP replaces global sorting with local probability mapping, reducing hardware complexity while suppressing herding and oscillations caused by stale states. Experiments on a cycle-accurate simulator show that PSP is robust across port scales, bandwidth-limited paths, and fixed-flow interference. It outperforms join-the-shortest-queue (JSQ) scheduling and Random in loss rate, 99th-percentile buffer occupancy, and scalability, while remaining competitive with Top-k at lower hardware cost. PSP provides an effective balance among performance, stability, and overhead for artificial intelligence data centers.

cs.NI↗

arXiv2025

Nanoscale Mapping of Transition Metal Ordering in Individual LiNi0.5Mn1.5O4 Particles Using 4D-STEM ACOM Technique

The electrochemical performance of the spinel LiNi0.5Mn1.5O4, a high-voltage positive electrode material for Li-ion batteries, is influenced by the transition metal arrangement in the octahedral network, leading to disordered (Fd m S.G.) and ordered3 (P4332 S.G.) structures. However, widely used techniques lack the spatial resolution necessary to elucidate the ordering phenomenon at the particle scale. Using the 4D-STEM technique, we present the first direct observation of ordering distribution in individual LiNi0.5Mn1.5O4 particles with nanometric spatial resolution. We propose a quantification method for the local degree of ordering based on the ratio of ordered to disordered spinel lattices along the particle thickness extracted from electron diffraction spot intensities. In an ordered spinel LiNi0.5Mn1.5O4, the transition metal ordering is consistently observed throughout the primary particle. However, the extent of ordering in the spinel phase depends on its distribution at the particle scale, a factor influenced by the annealing conditions. The 4D-STEM analysis elucidates the boundary between highly-ordered and low-ordered LiNi0.5Mn1.5O4 particles.

cond-mat.mtrl-sci↗

arXiv2025

Hierarchical and Step-Layer-Wise Tuning of Attention Specialty for Multi-Instance Synthesis in Diffusion Transformers

Text-to-image (T2I) generation models often struggle with multi-instance synthesis (MIS), where they must accurately depict multiple distinct instances in a single image based on complex prompts detailing individual features. Traditional MIS control methods for UNet architectures like SD v1.5/SDXL fail to adapt to DiT-based models like FLUX and SD v3.5, which rely on integrated attention between image and text tokens rather than text-image cross-attention. To enhance MIS in DiT, we first analyze the mixed attention mechanism in DiT. Our token-wise and layer-wise analysis of attention maps reveals a hierarchical response structure: instance tokens dominate early layers, background tokens in middle layers, and attribute tokens in later layers. Building on this observation, we propose a training-free approach for enhancing MIS in DiT-based models with hierarchical and step-layer-wise attention specialty tuning (AST). AST amplifies key regions while suppressing irrelevant areas in distinct attention maps across layers and steps, guided by the hierarchical structure. This optimizes multimodal interactions by hierarchically decoupling the complex prompts with instance-based sketches. We evaluate our approach using upgraded sketch-based layouts for the T2I-CompBench and customized complex scenes. Both quantitative and qualitative results confirm our method enhances complex layout generation, ensuring precise instance placement and attribute representation in MIS.

cs.CV↗

arXiv2025

Academic Network Representation via Prediction-Sampling Incorporated Tensor Factorization

Accurate representation to an academic network is of great significance to academic relationship mining like predicting scientific impact. A Latent Factorization of Tensors (LFT) model is one of the most effective models for learning the representation of a target network. However, an academic network is often High-Dimensional and Incomplete (HDI) because the relationships among numerous network entities are impossible to be fully explored, making it difficult for an LFT model to learn accurate representation of the academic network. To address this issue, this paper proposes a Prediction-sampling-based Latent Factorization of Tensors (PLFT) model with two ideas: 1) constructing a cascade LFT architecture to enhance model representation learning ability via learning academic network hierarchical features, and 2) introducing a nonlinear activation-incorporated predicting-sampling strategy to more accurately learn the network representation via generating new academic network data layer by layer. Experimental results from the three real-world academic network datasets show that the PLFT model outperforms existing models when predicting the unexplored relationships among network entities.

cs.LG↗

arXiv2023

SegMobaTree: The Segmented Multilayer Online Balanced Tree for high-performance IPv6 Lookup in the Edge Network

With the development of IPv6 and edge computing, the edge network should support IPv6 lookup (the longest prefix matching, LPM) with high lookup speed, high update speed, and low memory cost. However, the trie-based algorithms, e.g., SAIL and Poptrie, mainly focus on the IPv4 ruleset but have disadvantages in the edge IPv6 ruleset with longer prefix length. The binary-based algorithm Hi-BST also has limited lookup speed with too many memory accesses. Therefore, we propose the SegMobaTree algorithm to achieve high-performance IPv6 lookup in the edge network. First, MobaTree is a multilayer online balanced tree to perform high-speed binary search among rules with different prefix lengths. Second, to avoid one large tree, we propose the dynamic programming method to split prefix lengths into a few suitable segments, which tradeoff between the number of segments and the scale of trees. Compared to SAIL, Poptrie, and Hi-BST, SegMobaTree achieves 1.5x, 1.7x, 1.6x lookup speed and 11.8x, 32.8x, 1.3x update speed with low memory cost.

cs.NI↗

arXiv2023

An integral sliding-mode parallel control approach for general nonlinear systems via piecewise affine linear models

The fundamental problem of stabilizing a general nonaffine continuous-time nonlinear system is investigated via piecewise affine linear models (PALMs) in this article. A novel integral sliding-mode parallel control (ISMPC) approach is developed, where an uncertain piecewise affine system (PWA) is constructed to model a nonaffine continuous-time nonlinear system equivalently on a compact region containing the origin. A piecewise sliding-mode parallel controller is designed to globally stabilize the PALM and, consequently, to semiglobally stabilize the original nonlinear system. The proposed scheme enjoys three favorable features: (i) some restrictions on the system input channel are eliminated, thus the developed method is more relaxed compared with the published approaches; (ii) it is convenient to be used to deal with both matched and unmatched uncertainties of the system; and (iii) the proposed piecewise parallel controller generates smooth control signals even around the boundaries between different subspaces, which makes the developed control strategy more implementable and reliable. Moreover, we provide discussions about the universality analysis of the developed control strategy for two kinds of typical nonlinear systems. Simulation results from two numerical examples further demonstrate the performance of the developed control approach.

eess.SY↗

arXiv2021

Graph Representation Learning via Contrasting Cluster Assignments

With the rise of contrastive learning, unsupervised graph representation learning has been booming recently, even surpassing the supervised counterparts in some machine learning tasks. Most of existing contrastive models for graph representation learning either focus on maximizing mutual information between local and global embeddings, or primarily depend on contrasting embeddings at node level. However, they are still not exquisite enough to comprehensively explore the local and global views of network topology. Although the former considers local-global relationship, its coarse global information leads to grudging cooperation between local and global views. The latter pays attention to node-level feature alignment, so that the role of global view appears inconspicuous. To avoid falling into these two extreme cases, we propose a novel unsupervised graph representation model by contrasting cluster assignments, called as GRCCA. It is motivated to make good use of local and global information synthetically through combining clustering algorithms and contrastive learning. This not only facilitates the contrastive effect, but also provides the more high-quality graph information. Meanwhile, GRCCA further excavates cluster-level information, which make it get insight to the elusive association between nodes beyond graph topology. Specifically, we first generate two augmented graphs with distinct graph augmentation strategies, then employ clustering algorithms to obtain their cluster assignments and prototypes respectively. The proposed GRCCA further compels the identical nodes from different augmented graphs to recognize their cluster assignments mutually by minimizing a cross entropy loss. To demonstrate its effectiveness, we compare with the state-of-the-art models in three different downstream tasks. The experimental results show that GRCCA has strong competitiveness in most tasks.

cs.LG↗

arXiv2021

Microstructure and correlated mechanical properties study of Ni-(Fe, Co)-Mn-(Al, In) as-spun ribbons

The Ni-Mn-based shape memory alloys as a promising candidate of elastocaloric material has been reported in many literatures, especially on bulk samples. The as-spun ribbon, which has a larger surface area and is more efficient for heat transfer, is rarely studied and hence of importance. In the present work, we succeeded in producing very long as-spun Ni-Fe-Mn-(Al, In) ribbons, with around 300 mm in length. The microstructure and mechanical properties of these as-spun ribbons were thoroughly investigated by scanning electron microscopy / electron backscattered diffraction (SEM/EBSD), nanoindentation and 3-points bending experiments. Through SEM/EBSD analyses, the microstructure and texture of the as-spun ribbons were studied. A gradient in microstructure exists along the thickness direction (TD) of the ribbon, which is induced by the temperature gradient during fast rate solidification, resulting in fine equiaxed grains along the surface contacted with the rotating wheel in melt spinning process and elongated grains, respectively. Both equiaxed and elongated grains possess a strong {001} fiber texture (<001>//TD). Nanoindentation analyses show little variation of hardness between the two different microstructures. The ductility index of both Ni-Fe-Mn-Al and Ni-Fe-Mn-In ribbons are within the range of intermetallic materials. The substitution of In by Al allows to increase very slightly the ductility index, which can reach 0.75. The fine equiaxed grains show better tensile resistance than the elongated grains in 3-points bending test. The substitution of In by Al improves the maximum bending strain by a factor of 3. The maximum strain for Ni-Fe-Mn-Al as-spun ribbons can reach 3 % before fracture. Fractography shows that the intergranular fracture is the main damage mechanism in these as-spun ribbons.

cond-mat.mtrl-sci↗