서브메뉴
검색
Scaling Dataflow: Programmability and Simulation
Scaling Dataflow: Programmability and Simulation
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211153058
- ISBN
- 9798346381693
- DDC
- 005
- 저자명
- Zhang, Nathan.
- 서명/저자
- Scaling Dataflow: Programmability and Simulation
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 128 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-05, Section: A.
- 주기사항
- Advisor: Olukotun, Oyekunle.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약In recent years, machine learning has driven massively increased computational demand at all scales - ranging from kiloflop-scale per-packet networking analytics to petaflop-scale large language models (LLMs). Spatial accelerators such as field programmable gate arrays (FPGAs) and other reconfigurable dataflow architectures (RDAs) have emerged as promising candidates, offering significant performance-per-watt improvements over instruction-based architectures such as CPUs and GPUs. However, their widespread adoption is hampered by limitations in programming models and tools. This dissertation tackles two key challenges: (1) constructing efficiently composable libraries for these accelerators, and (2) enabling the parallel simulation of dataflow systems, which are distributed processing units connected by communication channels.Previous work in programming spatial accelerators made writing individual applications significantly easier by lifting the level of abstraction. However, writing single applications is insufficient; reusable libraries are essential to compose larger programs. We first characterize the hierarchical pipeline promotion problem (HPPP), a performance trap which prevents the creation of high performance libraries when implemented with traditional abstractions. We then present streaming tensor interfaces (STIs), a software design pattern which bypasses the HPPP by constructing decoupled streaming pipelines. Using STIs, we construct an efficient library for performing inference, achieving a throughput of 10 cycles-per-inference without batching.We then turn our focus to the simulation of dataflow systems, which sits at the heart of research - from functional testing to performance estimation. Unfortunately, existing parallel software simulation schemes are ill-suited to such systems; optimistic schemes place a large burden on the end-user to reason about undoing speculation errors, while current conservative schemes scale poorly in the face of low-latency high-bandwidth communication. To address these limitations, we introduce the Dataflow Abstract Machine (DAM), a parallel simulator framework which breaks from tradition in both user interface and execution. DAM replaces event-driven modeling with communicating sequential processes, bulk-synchronous time with asynchronous distributed time, and global synchronization with a peer-to-peer protocol. DAM requires 57% less code compared to an existing cycle-based simulation, and outperforms a state-of-the-art simulation framework by at least 2x. Compared to ad-hoc research simulators, DAM achieves speedups of up to five orders of magnitude.Ultimately, both of these advancements are enabled by new abstractions. In this thesis, we will argue that abstractions are key, as abstractions constrain implementations, and implementations dictate performance.
- 일반주제명
- Programming languages
- 일반주제명
- Systems design
- 일반주제명
- Libraries
- 일반주제명
- Software engineering
- 일반주제명
- Linear algebra
- 일반주제명
- Semantics
- 일반주제명
- Engineers
- 일반주제명
- Computer engineering
- 일반주제명
- Computer science
- 일반주제명
- Design
- 일반주제명
- Logic
- 일반주제명
- Systems science
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-05A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164878
■00520250211153058
■006m o d
■007cr#unu||||||||
■020 ▼a9798346381693
■035 ▼a(MiAaPQ)AAI31652036
■035 ▼a(MiAaPQ)Stanfordbc169pf3686
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a005
■1001 ▼aZhang, Nathan.
■24510▼aScaling Dataflow: Programmability and Simulation
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a128 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-05, Section: A.
■500 ▼aAdvisor: Olukotun, Oyekunle.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aIn recent years, machine learning has driven massively increased computational demand at all scales - ranging from kiloflop-scale per-packet networking analytics to petaflop-scale large language models (LLMs). Spatial accelerators such as field programmable gate arrays (FPGAs) and other reconfigurable dataflow architectures (RDAs) have emerged as promising candidates, offering significant performance-per-watt improvements over instruction-based architectures such as CPUs and GPUs. However, their widespread adoption is hampered by limitations in programming models and tools. This dissertation tackles two key challenges: (1) constructing efficiently composable libraries for these accelerators, and (2) enabling the parallel simulation of dataflow systems, which are distributed processing units connected by communication channels.Previous work in programming spatial accelerators made writing individual applications significantly easier by lifting the level of abstraction. However, writing single applications is insufficient; reusable libraries are essential to compose larger programs. We first characterize the hierarchical pipeline promotion problem (HPPP), a performance trap which prevents the creation of high performance libraries when implemented with traditional abstractions. We then present streaming tensor interfaces (STIs), a software design pattern which bypasses the HPPP by constructing decoupled streaming pipelines. Using STIs, we construct an efficient library for performing inference, achieving a throughput of 10 cycles-per-inference without batching.We then turn our focus to the simulation of dataflow systems, which sits at the heart of research - from functional testing to performance estimation. Unfortunately, existing parallel software simulation schemes are ill-suited to such systems; optimistic schemes place a large burden on the end-user to reason about undoing speculation errors, while current conservative schemes scale poorly in the face of low-latency high-bandwidth communication. To address these limitations, we introduce the Dataflow Abstract Machine (DAM), a parallel simulator framework which breaks from tradition in both user interface and execution. DAM replaces event-driven modeling with communicating sequential processes, bulk-synchronous time with asynchronous distributed time, and global synchronization with a peer-to-peer protocol. DAM requires 57% less code compared to an existing cycle-based simulation, and outperforms a state-of-the-art simulation framework by at least 2x. Compared to ad-hoc research simulators, DAM achieves speedups of up to five orders of magnitude.Ultimately, both of these advancements are enabled by new abstractions. In this thesis, we will argue that abstractions are key, as abstractions constrain implementations, and implementations dictate performance.
■590 ▼aSchool code: 0212.
■650 4▼aProgramming languages
■650 4▼aSystems design
■650 4▼aLibraries
■650 4▼aSoftware engineering
■650 4▼aLinear algebra
■650 4▼aSemantics
■650 4▼aEngineers
■650 4▼aComputer engineering
■650 4▼aComputer science
■650 4▼aDesign
■650 4▼aLogic
■650 4▼aSystems science
■690 ▼a0800
■690 ▼a0464
■690 ▼a0984
■690 ▼a0389
■690 ▼a0501
■690 ▼a0395
■690 ▼a0790
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-05A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164878▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


