서브메뉴
검색
Efficiently Scaling Machine Learning Systems Across Heterogeneous Resources
Efficiently Scaling Machine Learning Systems Across Heterogeneous Resources
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105245
- ISBN
- 9798291569672
- DDC
- 621.3
- 저자명
- Jin, Shuowei.
- 서명/저자
- Efficiently Scaling Machine Learning Systems Across Heterogeneous Resources
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 116 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Mao, Z. Morley.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약The rapid advancement of machine learning (ML), particularly deep learning and Large Language Models (LLMs), has revolutionized numerous domains. This progress is driven by scaling laws, indicating that larger models trained on more data with more compute yield better performance. However, this scaling introduces significant challenges, especially during inference, due to massive model sizes and computational demands. Furthermore, emerging inference-time scaling laws, where allocating more compute during inference boosts performance, add another layer of complexity. Efficiently scaling ML inference systems to optimize critical metrics like latency and throughput is further complicated by the inherent heterogeneity of computing resources, spanning from resource-constrained edge devices to diverse hardware within datacenters. Naive scaling approaches often lead to suboptimal performance and resource utilization in these heterogeneous environments.This dissertation argues that efficiently scaling ML systems across heterogeneous resources requires a holistic approach. We adopt Algorithm-System Co-design as the central methodology, simultaneously optimizing ML algorithms and the underlying system components (hardware, software, schedulers) by considering their intricate interactions. This co-design philosophy allows us to overcome the limitations of optimizing algorithms and systems in isolation, unlocking significant efficiency gains.We demonstrate the efficacy of this approach through three core contributions. First, OASIS, a system for collaborative neural-enhanced mobile video streaming. It addresses the challenge of running demanding super-resolution models on resource-constrained mobile devices with varying network conditions. OASIS co-designs adaptive bitrate/model selection algorithms with a cross-device collaborative inference mechanism, enabling multiple nearby devices to efficiently share computation resources, significantly improving video Quality of Experience (QoE) and reducing energy consumption. Second, Cake, a system optimizing Large Language Model (LLM) inference latency, specifically targeting the Key-Value (KV) cache generation bottleneck during long-context prefill. It tackles the compute vs. I/O trade-off in heterogeneous memory systems by co-designing a bidirectional parallel scheduling strategy that dynamically balances computing the KV cache on GPUs and loading it from storage, substantially reducing Time-To-First-Token (TTFT). Third, Plato, a semantic-aware parallel decoding framework for LLMs that addresses the inefficiency of autoregressive decoding while maintaining answer quality. Plato co-designs algorithmic components (organizing sub-problems into dependency graphs based on logical relationships) with system optimizations (pipelined execution and KV cache reuse). Our evaluations show that Plato improves throughput by up to 68% over autoregressive decoding while maintaining a 90% quality net-win rate compared to other parallel decoding methods like Skeleton-of-Thought.Collectively, these contributions illustrate that principled algorithm-system co-design is a powerful and versatile methodology for building efficient ML inference systems that can effectively scale across diverse and heterogeneous resources, optimizing for latency and throughput from the edge to the datacenter.
- 일반주제명
- Computer engineering
- 일반주제명
- Computer science
- 기타저자
- University of Michigan Computer Science & Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359981
■00520260202105245
■006m o d
■007cr#unu||||||||
■020 ▼a9798291569672
■035 ▼a(MiAaPQ)AAI32272042
■035 ▼a(MiAaPQ)umichrackham006341
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aJin, Shuowei.
■24510▼aEfficiently Scaling Machine Learning Systems Across Heterogeneous Resources
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a116 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Mao, Z. Morley.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aThe rapid advancement of machine learning (ML), particularly deep learning and Large Language Models (LLMs), has revolutionized numerous domains. This progress is driven by scaling laws, indicating that larger models trained on more data with more compute yield better performance. However, this scaling introduces significant challenges, especially during inference, due to massive model sizes and computational demands. Furthermore, emerging inference-time scaling laws, where allocating more compute during inference boosts performance, add another layer of complexity. Efficiently scaling ML inference systems to optimize critical metrics like latency and throughput is further complicated by the inherent heterogeneity of computing resources, spanning from resource-constrained edge devices to diverse hardware within datacenters. Naive scaling approaches often lead to suboptimal performance and resource utilization in these heterogeneous environments.This dissertation argues that efficiently scaling ML systems across heterogeneous resources requires a holistic approach. We adopt Algorithm-System Co-design as the central methodology, simultaneously optimizing ML algorithms and the underlying system components (hardware, software, schedulers) by considering their intricate interactions. This co-design philosophy allows us to overcome the limitations of optimizing algorithms and systems in isolation, unlocking significant efficiency gains.We demonstrate the efficacy of this approach through three core contributions. First, OASIS, a system for collaborative neural-enhanced mobile video streaming. It addresses the challenge of running demanding super-resolution models on resource-constrained mobile devices with varying network conditions. OASIS co-designs adaptive bitrate/model selection algorithms with a cross-device collaborative inference mechanism, enabling multiple nearby devices to efficiently share computation resources, significantly improving video Quality of Experience (QoE) and reducing energy consumption. Second, Cake, a system optimizing Large Language Model (LLM) inference latency, specifically targeting the Key-Value (KV) cache generation bottleneck during long-context prefill. It tackles the compute vs. I/O trade-off in heterogeneous memory systems by co-designing a bidirectional parallel scheduling strategy that dynamically balances computing the KV cache on GPUs and loading it from storage, substantially reducing Time-To-First-Token (TTFT). Third, Plato, a semantic-aware parallel decoding framework for LLMs that addresses the inefficiency of autoregressive decoding while maintaining answer quality. Plato co-designs algorithmic components (organizing sub-problems into dependency graphs based on logical relationships) with system optimizations (pipelined execution and KV cache reuse). Our evaluations show that Plato improves throughput by up to 68% over autoregressive decoding while maintaining a 90% quality net-win rate compared to other parallel decoding methods like Skeleton-of-Thought.Collectively, these contributions illustrate that principled algorithm-system co-design is a powerful and versatile methodology for building efficient ML inference systems that can effectively scale across diverse and heterogeneous resources, optimizing for latency and throughput from the edge to the datacenter.
■590 ▼aSchool code: 0127.
■650 4▼aComputer engineering
■650 4▼aComputer science
■653 ▼aMachine learning systems
■653 ▼aLarge Language Models
■653 ▼aEnergy consumption
■690 ▼a0984
■690 ▼a0464
■690 ▼a0800
■71020▼aUniversity of Michigan▼bComputer Science & Engineering.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359981▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


