서브메뉴
검색
Systems, Models and Algorithms for Failure Diagnosis in Networked Infrastructure
Systems, Models and Algorithms for Failure Diagnosis in Networked Infrastructure
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105707
- ISBN
- 9798265456168
- DDC
- 004
- 저자명
- Harsh, Vipul.
- 서명/저자
- Systems, Models and Algorithms for Failure Diagnosis in Networked Infrastructure
- 발행사항
- [Sl] : University of Illinois at Urbana-Champaign, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 113 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-06, Section: A.
- 주기사항
- Advisor: Godfrey, P. Brighten.
- 학위논문주기
- Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2024.
- 초록/해제
- 요약Failure incidents in networked systems such as datacenters, private or public cloud environments, WAN and enterprise networks lead to significant downtime and violations of service-level agreements, incurring large financial losses. To mitigate failures quickly, operators seek to implement automated failure diagnosis, also known as Root Cause Analysis (RCA). The goal of RCA is to localize faults with high accuracy in a timely manner using practical monitoring telemetry. There are two major challenges in designing such RCA solutions: (1) accurately modeling the behaviour of the system and (2) using the model to infer the root causes accurately. Existing works on RCA either (1) fall short in modeling the complexities of an environment and hence utilize imprecise models or (2) utilize an inference algorithm for RCA that is inadequate in extracting high accuracy from the model.In this thesis, we propose solutions for RCA based on two high level ideas. First, we utilize a suitably designed Probabilistic Graphical Model (PGM) to accurately model the system behaviour. A PGM is a model based on an underlying graph structure that, as we show in the thesis, naturally allows us to model the complexities of a distributed system with many components. Second, we design custom training and inference algorithms to extract maximum inference accuracy from the PGM-based model. Our algorithms are designed to handle the various characteristics of real world systems such as large scale or noisy input monitoring data.Using the high level ideas outlined above, we propose three RCA systems. In chapter 2, we describe Murphy, a RCA system for diagnosing performance issues in distributed applications. Murphy models weak inter-dependencies between system components via a Markov Random Field (a type of PGM based on undirected graphs) that can handle cyclic dependencies. With its more precise modeling of inter-dependencies, it is able to predict root causes with higher accuracy than past works. In chapter 3, we describe Flock that localizes unreported failures in datacenter networks. Flock models the problem via a Bayesian network (a type of PGM based on directed acyclic graphs) that can handle the various complexities of a datacenter network such as lack of knowledge of path information of flows or occasional noisy packet drops. Flock utilizes a custom inference algorithm that can speed up inference in discrete-valued PGMs by multiple orders of magnitude. This inference algorithm allows Flock to unlock the benefits of PGM-based modeling for datacenter networks and allows it to achieve higher inference accuracy than past works. In chapter 4, we describe FaultFerence, a RCA system for detecting faults in datacenter networks with only passive monitoring that obscures path information up to a large set of paths, for all flows. In such cases, even Flock (or any fault localization algorithm) can only localize the fault to a large equivalent set of devices. To localize within this equivalence set, FaultFerence uses a similar model as Flock's but additionally runs iterations of inference followed by strategically chosen "microactions" that tweak the network ever so slightly which helps in symmetry breaking. Using this approach, FaultFerence is able to reduce the time to localize faults, while also reducing the invasiveness of the localization procedure compared to ad-hoc techniques employed today. The above systems demonstrate the benefits of principled inference via accurate modeling of system behaviour towards achieving high diagnosis accuracy.
- 일반주제명
- Computer science
- 일반주제명
- Statistics
- 일반주제명
- Communication
- 일반주제명
- Information technology
- 기타저자
- University of Illinois at Urbana-Champaign Computer Science
- 기본자료저록
- Dissertations Abstracts International. 87-06A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2024 us c eng d■001000017361107
■00520260202105707
■006m o d
■007cr#unu||||||||
■020 ▼a9798265456168
■035 ▼a(MiAaPQ)AAI32435152
■035 ▼a(MiAaPQ)124409
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aHarsh, Vipul.
■24510▼aSystems, Models and Algorithms for Failure Diagnosis in Networked Infrastructure
■260 ▼a[Sl]▼bUniversity of Illinois at Urbana-Champaign▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a113 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-06, Section: A.
■500 ▼aAdvisor: Godfrey, P. Brighten.
■5021 ▼aThesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2024.
■520 ▼aFailure incidents in networked systems such as datacenters, private or public cloud environments, WAN and enterprise networks lead to significant downtime and violations of service-level agreements, incurring large financial losses. To mitigate failures quickly, operators seek to implement automated failure diagnosis, also known as Root Cause Analysis (RCA). The goal of RCA is to localize faults with high accuracy in a timely manner using practical monitoring telemetry. There are two major challenges in designing such RCA solutions: (1) accurately modeling the behaviour of the system and (2) using the model to infer the root causes accurately. Existing works on RCA either (1) fall short in modeling the complexities of an environment and hence utilize imprecise models or (2) utilize an inference algorithm for RCA that is inadequate in extracting high accuracy from the model.In this thesis, we propose solutions for RCA based on two high level ideas. First, we utilize a suitably designed Probabilistic Graphical Model (PGM) to accurately model the system behaviour. A PGM is a model based on an underlying graph structure that, as we show in the thesis, naturally allows us to model the complexities of a distributed system with many components. Second, we design custom training and inference algorithms to extract maximum inference accuracy from the PGM-based model. Our algorithms are designed to handle the various characteristics of real world systems such as large scale or noisy input monitoring data.Using the high level ideas outlined above, we propose three RCA systems. In chapter 2, we describe Murphy, a RCA system for diagnosing performance issues in distributed applications. Murphy models weak inter-dependencies between system components via a Markov Random Field (a type of PGM based on undirected graphs) that can handle cyclic dependencies. With its more precise modeling of inter-dependencies, it is able to predict root causes with higher accuracy than past works. In chapter 3, we describe Flock that localizes unreported failures in datacenter networks. Flock models the problem via a Bayesian network (a type of PGM based on directed acyclic graphs) that can handle the various complexities of a datacenter network such as lack of knowledge of path information of flows or occasional noisy packet drops. Flock utilizes a custom inference algorithm that can speed up inference in discrete-valued PGMs by multiple orders of magnitude. This inference algorithm allows Flock to unlock the benefits of PGM-based modeling for datacenter networks and allows it to achieve higher inference accuracy than past works. In chapter 4, we describe FaultFerence, a RCA system for detecting faults in datacenter networks with only passive monitoring that obscures path information up to a large set of paths, for all flows. In such cases, even Flock (or any fault localization algorithm) can only localize the fault to a large equivalent set of devices. To localize within this equivalence set, FaultFerence uses a similar model as Flock's but additionally runs iterations of inference followed by strategically chosen "microactions" that tweak the network ever so slightly which helps in symmetry breaking. Using this approach, FaultFerence is able to reduce the time to localize faults, while also reducing the invasiveness of the localization procedure compared to ad-hoc techniques employed today. The above systems demonstrate the benefits of principled inference via accurate modeling of system behaviour towards achieving high diagnosis accuracy.
■590 ▼aSchool code: 0090.
■650 4▼aComputer science
■650 4▼aStatistics
■650 4▼aCommunication
■650 4▼aInformation technology
■653 ▼aPerformance diagnosis
■653 ▼aFault localization
■653 ▼aRoot Cause Analysis
■653 ▼aProbabilistic graphical models
■653 ▼aBayesian networks
■690 ▼a0984
■690 ▼a0489
■690 ▼a0459
■690 ▼a0463
■71020▼aUniversity of Illinois at Urbana-Champaign▼bComputer Science.
■7730 ▼tDissertations Abstracts International▼g87-06A.
■790 ▼a0090
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17361107▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


