Difference between revisions of "Resource:Seminar"

From MobiNetS
Jump to: navigation, search
Line 8: Line 8:


{{Latest_seminar
{{Latest_seminar
|abstract = As Large Language Models (LLMs) continue to scale, optimizing their deployment requires efficient hardware and system co-design. However, current LLM performance evaluation frameworks fail to capture both chip-level execution details and system-wide behavior, making it difficult to assess realistic performance bottlenecks. In this work, we introduce ReaLLM, a trace-driven simulation framework designed to bridge the gap between detailed accelerator design and large-scale inference evaluation. Unlike prior simulators, ReaLLM integrates kernel profiling derived from detailed microarchitectural simulations with a new trace-driven end-to-end system simulator, enabling precise evaluation of parallelism strategies, batching techniques, and scheduling policies. To address the high computational cost of exhaustive simulations, ReaLLM constructs a precomputed kernel library based on hypothesized scenarios, interpolating results to efficiently explore a vast design space of LLM inference systems. Our validation against real hardware demonstrates the framework's accuracy, achieving an average end-to-end latency prediction error of only 9.1% when simulating inference tasks running on 4 NVIDIA H100 GPUs. We further use ReaLLM to evaluate popular LLMs' end-to-end performance across traces from different applications and identify key system bottlenecks, showing that modern GPU-based LLM inference is increasingly compute-bound rather than memory-bandwidth bound at large scale. Additionally, we significantly reduce simulation time with our precomputed kernel library by a factor of 6× for full-simulations and 164× for workload SLO exploration. ReaLLM is open-source and available at https://github.com/bespoke-silicon-group/reallm..
|abstract = Running deep neural networks (DNNs) on large-scale videos from widely distributed cameras presents two significant challenges. Firstly, video quality for analytical purposes is severely impacted by the camera deployment environment, which is termed Pixel Recession in this paper. Secondly, low-latency video streaming from the source camera to edge servers is greatly hindered by the rapid expansion of video traffic. Despite numerous efforts such as enhancing the video structure, uneven encoding, and filtering frames captured on camera, these methods have proven insufficient to address the challenges at hand. We propose Spliceosome, a novel video analytics system that effectively overcomes the pixel recession and streaming bottlenecks. In brief, Spliceosome 1) recovers from pixel recession by adaptive video knobs (i.e., brightness and contrast) tuning in ARP (anchor region proposal) granularity, and 2) lowers the transmission volume by video thinning, which uses only single-channel information for video encoding. We implemented Spliceosome using only commercial off-the-shelf hardware. Our experimental results demonstrate that Spliceosome outperforms other alternative designs by 4.71-14.47%, 40.94-58.71%, and 14.28% in detection accuracy, end-to-end delay, and efficiency of DNNs inference, respectively.
|confname =ASAP'25
|confname =ToN'25
|link = https://ieeexplore.ieee.org/abstract/document/11113621
|link = https://ieeexplore.ieee.org/abstract/document/10843977
|title= ReaLLM: A Trace-Driven Framework for Rapid Simulation of Large-Scale LLM Inference
|title= Spliceosome: On-Camera Video Thinning and Tuning for Timely and Accurate Analytics
|speaker=JunZhe
|speaker=Zhongwei Sun
|date=2025-11-21
|date=2025-11-28
}}{{Latest_seminar
}}{{Latest_seminar
|abstract =With the proliferation of mobile devices, spatial crowdsourcing has emerged as a promising paradigm for facilitating location-based services, encompassing various applications across academia and industries. Recently, pioneering works have attempted to infer workers' mobility patterns from historical data to improve the quality of task assignment. However, these studies have overlooked or under-examined issues such as the dynamic mobility patterns of crowd workers, especially in the context of newcomers, the misalignment between the objectives of mobility prediction and task assignment, and the effective utilization of predicted mobility patterns. In this paper, we investigate a problem we term Task Assignment in Mobility Prediction-aware Spatial Crowdsourcing (TAMP). To address the TAMP problem, we first propose a task-adaptive meta-learning algorithm, which trains a set of specific meta-knowledge for workers' mobility prediction models through game theory-based learning task clustering and meta-training within each cluster. Then, we design a task assignment-oriented loss function and develop a task assignment algorithm that incorporates prediction performance, prioritizing assignments with higher confidence of completion. Extensive experiments on real-world datasets validate that our proposed methods can effectively improve the quality of task assignment.
|abstract =The rapid expansion of large language models (LLMs) requires the development of extensive GPU clusters, with companies deploying clusters with tens to hundreds of thousands of GPUs. This growth significantly expands the design space for LLM training systems, requiring thorough exploration of different parallelization strategies, communication parameters, congestion control, fabric topology, etc. Current methods require up to 10k simulation experiments to identify optimal configurations, with inadequate exploration leading to significant degradation of training performance. In this paper, we tackle the overlooked problem of efficiently conducting parallel simulation experiments for design space exploration. Our analysis and experiments show that Single-process Multi-experiment (SPME) achieves superior performance by reducing scheduling overhead and optimizing resource utilization, yet remains insufficient for current AI cluster scales. To enhance SPME’s efficacy, we introduce Multiverse, a novel GPU-based AI training simulator. Multiverse leverages the computing throughput of GPUs efficiently with optimizations such as a pull-based synchronization, highfidelity intra-server communication, and a kernel-fusion technique. Extensive experiments validate the accuracy and efficiency of Multiverse, demonstrating less than 3.0% discrepancy with real-world LLM training on clusters of up to 54,000 GPUs, achieving 43.1−73.2X speedup over state-of-the-art CPU-based simulators in various use cases.
|confname =ICDE'25
|confname =NSDI'25
|link = https://ieeexplore.ieee.org/document/11113007
|link = https://www.usenix.org/conference/nsdi25/presentation/gui
|title= Effective Task Assignment in Mobility Prediction-Aware Spatial Crowdsourcing
|title= Accelerating Design Space Exploration for LLM Training Systems with Multi-experiment Parallel Simulation
|speaker= Zhenguo
|speaker=Qinyong
|date=2025-11-21
|date=2025-11-28
}}
}}
{{Resource:Previous_Seminars}}
{{Resource:Previous_Seminars}}

Revision as of 05:38, 28 November 2025

Time: 2025-11-21 10:30
Address: 4th Research Building A518
Useful links: 📚 Readling list; 📆 Schedules; 🧐 Previous seminars.

Latest

  1. [ToN'25] Spliceosome: On-Camera Video Thinning and Tuning for Timely and Accurate Analytics, Zhongwei Sun
    Abstract: Running deep neural networks (DNNs) on large-scale videos from widely distributed cameras presents two significant challenges. Firstly, video quality for analytical purposes is severely impacted by the camera deployment environment, which is termed Pixel Recession in this paper. Secondly, low-latency video streaming from the source camera to edge servers is greatly hindered by the rapid expansion of video traffic. Despite numerous efforts such as enhancing the video structure, uneven encoding, and filtering frames captured on camera, these methods have proven insufficient to address the challenges at hand. We propose Spliceosome, a novel video analytics system that effectively overcomes the pixel recession and streaming bottlenecks. In brief, Spliceosome 1) recovers from pixel recession by adaptive video knobs (i.e., brightness and contrast) tuning in ARP (anchor region proposal) granularity, and 2) lowers the transmission volume by video thinning, which uses only single-channel information for video encoding. We implemented Spliceosome using only commercial off-the-shelf hardware. Our experimental results demonstrate that Spliceosome outperforms other alternative designs by 4.71-14.47%, 40.94-58.71%, and 14.28% in detection accuracy, end-to-end delay, and efficiency of DNNs inference, respectively.
  2. [NSDI'25] Accelerating Design Space Exploration for LLM Training Systems with Multi-experiment Parallel Simulation, Qinyong
    Abstract: The rapid expansion of large language models (LLMs) requires the development of extensive GPU clusters, with companies deploying clusters with tens to hundreds of thousands of GPUs. This growth significantly expands the design space for LLM training systems, requiring thorough exploration of different parallelization strategies, communication parameters, congestion control, fabric topology, etc. Current methods require up to 10k simulation experiments to identify optimal configurations, with inadequate exploration leading to significant degradation of training performance. In this paper, we tackle the overlooked problem of efficiently conducting parallel simulation experiments for design space exploration. Our analysis and experiments show that Single-process Multi-experiment (SPME) achieves superior performance by reducing scheduling overhead and optimizing resource utilization, yet remains insufficient for current AI cluster scales. To enhance SPME’s efficacy, we introduce Multiverse, a novel GPU-based AI training simulator. Multiverse leverages the computing throughput of GPUs efficiently with optimizations such as a pull-based synchronization, highfidelity intra-server communication, and a kernel-fusion technique. Extensive experiments validate the accuracy and efficiency of Multiverse, demonstrating less than 3.0% discrepancy with real-world LLM training on clusters of up to 54,000 GPUs, achieving 43.1−73.2X speedup over state-of-the-art CPU-based simulators in various use cases.

History

|abstract =The rapid expansion of large language models (LLMs) requires the development of extensive GPU clusters, with companies deploying clusters with tens to hundreds of thousands of GPUs. This growth significantly expands the design space for LLM training systems, requiring thorough exploration of different parallelization strategies, communication parameters, congestion control, fabric topology, etc. Current methods require up to 10k simulation experiments to identify optimal configurations, with inadequate exploration leading to significant degradation of training performance. In this paper, we tackle the overlooked problem of efficiently conducting parallel simulation experiments for design space exploration. Our

2024

2023

2022

2021

2020

  • [Topic] [ The path planning algorithm for multiple mobile edge servers in EdgeGO], Rong Cong, 2020-11-18

2019

2018

2017

Instructions

请使用Latest_seminar和Hist_seminar模板更新本页信息.

    • 修改时间和地点信息
    • 将当前latest seminar部分的code复制到这个页面中
    • 将{{Latest_seminar... 修改为 {{Hist_seminar...，并增加对应的日期信息|date=
    • 填入latest seminar各字段信息
    • link请务必不要留空，如果没有link则填本页地址 https://mobinets.org/index.php?title=Resource:Seminar
  • 格式说明
    • Latest_seminar:

{{Latest_seminar
|confname=
|link=
|title=
|speaker=
}}

    • Hist_seminar

{{Hist_seminar
|confname=
|link=
|title=
|speaker=
|date=
}}

Retrieved from "http://mobinets.cn/site/index.php?title=Resource:Seminar&oldid=3449"