arXiv subjects

Biao Sun

Publications and source records attributed to Biao Sun.

8 recordsLinked to original sources

arXiv2026

Frequency Response of Windowed DFT Phasor Estimation: Impact on Oscillation Observability

Phasor measurement units (PMUs) are widely used for sub-synchronous oscillation monitoring, yet the effect of windowed discrete Fourier transform (DFT)-based phasor estimation on oscillation observability is not fully characterized. This letter derives the complete complex-valued frequency response of the windowed DFT phasor estimator under both magnitude and phase modulation. The analysis shows that the estimation window introduces both frequency-dependent magnitude attenuation and phase shift to oscillation components, governed by the complex gain. A simple recovery method is also proposed to restore the true oscillation amplitude and phase from PMU data using the analytically known complex gain. The results are validated through time-domain simulations and provide guidance for industry practitioners on interpreting PMU-based oscillation measurements and selecting appropriate window lengths.

eess.SP↗

arXiv2024

Llumnix: Dynamic Scheduling for Large Language Model Serving

Inference serving for large language models (LLMs) is the key to unleashing their potential in people's daily lives. However, efficient LLM serving remains challenging today because the requests are inherently heterogeneous and unpredictable in terms of resource and latency requirements, as a result of the diverse applications and the dynamic execution nature of LLMs. Existing systems are fundamentally limited in handling these characteristics and cause problems such as severe queuing delays, poor tail latencies, and SLO violations. We introduce Llumnix, an LLM serving system that reacts to such heterogeneous and unpredictable requests by runtime rescheduling across multiple model instances. Similar to context switching across CPU cores in modern operating systems, Llumnix reschedules requests to improve load balancing and isolation, mitigate resource fragmentation, and differentiate request priorities and SLOs. Llumnix implements the rescheduling with an efficient and scalable live migration mechanism for requests and their in-memory states, and exploits it in a dynamic scheduling policy that unifies the multiple rescheduling scenarios elegantly. Our evaluations show that Llumnix improves tail latencies by an order of magnitude, accelerates high-priority requests by up to 1.5x, and delivers up to 36% cost savings while achieving similar tail latencies, compared against state-of-the-art LLM serving systems. Llumnix is publicly available at https://github.com/AlibabaPAI/llumnix.

cs.AR↗

arXiv2022

EasyScale: Accuracy-consistent Elastic Training for Deep Learning

Distributed synchronized GPU training is commonly used for deep learning. The resource constraint of using a fixed number of GPUs makes large-scale training jobs suffer from long queuing time for resource allocation, and lowers the cluster utilization. Adapting to resource elasticity can alleviate this but often introduces inconsistent model accuracy, due to lacking of capability to decouple model training procedure from resource allocation. We propose EasyScale, an elastic training system that achieves consistent model accuracy under resource elasticity for both homogeneous and heterogeneous GPUs. EasyScale preserves the data-parallel training behaviors strictly, traces the consistency-relevant factors carefully, utilizes the deep learning characteristics for EasyScaleThread abstraction and fast context-switching. To utilize heterogeneous cluster, EasyScale dynamically assigns workers based on the intra-/inter-job schedulers, minimizing load imbalance and maximizing aggregated job throughput. Deployed in an online serving cluster, EasyScale powers the training jobs to utilize idle GPUs opportunistically, improving overall cluster utilization by 62.1%.

cs.DC↗

arXiv2022

An Olfactory EEG Signal Classification Network Based on Frequency Band Feature Extraction

Classification of olfactory-induced electroencephalogram (EEG) signals has shown great potential in many fields. Since different frequency bands within the EEG signals contain different information, extracting specific frequency bands for classification performance is important. Moreover, due to the large inter-subject variability of the EEG signals, extracting frequency bands with subject-specific information rather than general information is crucial. Considering these, the focus of this letter is to classify the olfactory EEG signals by exploiting the spectral-domain information of specific frequency bands. In this letter, we present an olfactory EEG signal classification network based on frequency band feature extraction. A frequency band generator is first designed to extract frequency bands via the sliding window technique. Then, a frequency band attention mechanism is proposed to optimize frequency bands for a specific subject adaptively. Last, a convolutional neural network (CNN) is constructed to extract the spatio-spectral information and predict the EEG category. Comparison experiment results reveal that the proposed method outperforms a series of baseline methods in terms of both classification quality and inter-subject robustness. Ablation experiment results demonstrate the effectiveness of each component of the proposed method.

eess.SP↗

arXiv2018

A Detection and Segmentation Architecture for Skin Lesion Segmentation on Dermoscopy Images

This report summarises our method and validation results for the ISIC Challenge 2018 - Skin Lesion Analysis Towards Melanoma Detection - Task 1: Lesion Segmentation. We present a two-stage method for lesion segmentation with optimised training method and ensemble post-process. Our method achieves state-of-the-art performance on lesion segmentation and we win the first place in ISIC 2018 task1.

cs.CV↗

arXiv2016

HISTORY: An Efficient and Robust Algorithm for Noisy 1-bit Compressed Sensing

We consider the problem of sparse signal recovery from 1-bit measurements. Due to the noise present in the acquisition and transmission process, some quantized bits may be flipped to their opposite states. These sign flips may result in severe performance degradation. In this study, a novel algorithm, termed HISTORY, is proposed. It consists of Hamming support detection and coefficients recovery. The HISTORY algorithm has high recovery accuracy and is robust to strong measurement noise. Numerical results are provided to demonstrate the effectiveness and superiority of the proposed algorithm.

cs.IT↗

arXiv2016

Compressed Sensing for Implantable Neural Recordings Using Co-sparse Analysis Model and Weighted $\ell_1$-Optimization

Reliable and energy-efficient wireless data transmission remains a major challenge in resource-constrained wireless neural recording tasks, where data compression is generally adopted to relax the burdens on the wireless data link. Recently, Compressed Sensing (CS) theory has successfully demonstrated its potential in neural recording application. The main limitation of CS, however, is that the neural signals have no good sparse representation with commonly used dictionaries and learning a reliable dictionary is often data dependent and computationally demanding. In this paper, a novel CS approach for implantable neural recording is proposed. The main contributions are: 1) The co-sparse analysis model is adopted to enforce co-sparsity of the neural signals, therefore overcoming the drawbacks of conventional synthesis model and enhancing the reconstruction performance. 2) A multi-fractional-order difference matrix is constructed as the analysis dictionary, thus avoiding the dictionary learning procedure and reducing the need for previously acquired data and computational resources. 3) By exploiting the statistical priors of the analysis coefficients, a weighted analysis $\ell_1$-minimization (WALM) algorithm is proposed to reconstruct the neural signals. Experimental results on Leicester neural signal database reveal that the proposed approach outperforms the state-of-the-art CS-based methods. On the challenging high compression ratio task, the proposed approach still achieves high reconstruction performance and spike classification accuracy.

cs.IT↗

arXiv2013

Physics design for the C-ADS main linac basing on two different injector design schemes

The China ADS (C-ADS) project is proposed to build a 1000 MW Accelerator Driven sub-critical System around 2032. The accelerator will work in CW mode with 10 mA in beam current and 1.5 GeV in final beam energy. The linac is composed of two major sections: the injector section and the main linac section. There are two different schemes for the injector section. The Injector-I scheme is based on a 325-MHz RFQ and superconducting spoke cavities of same RF frequency and the Injector-II scheme is based on a 162.5-MHz RFQ and superconducting HWR cavities of same frequency. The main linac design will be different for different injector choice. The two different designs for the main linac have been studied according to the beam characteristics from the different injector schemes.

physics.acc-ph↗