속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
l
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_f0ce02c2. 확인 범위: metadata_only. 확인한 자료: https://cslab.yonsei.ac.kr/publications. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_e0205794. 확인 범위: partial_source. 세부 범위: official_abstract_partial. 확인한 자료: https://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE08763759 ; https://cslab.yonsei.ac.kr/publications. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility. +
light dedup a light weight inline deduplication framework for non volatile memory file systems 000ae22e +
Verification: abstract_only; confidence=medium.
Canonical title: Light-Dedup: A Light-weight Inline Deduplication Framework for Non-Volatile Memory File Systems
Question: NVM 파일시스템에서 inline dedup의 메타데이터·쓰기 증폭을 줄일 수 있는가?
Context: 전역 fingerprint index는 NVM 접근과 metadata amplification이 크다.
Method: 지역성 인지 region buffer/index와 NVM 내 metadata table, prefetch를 사용한다.
Evaluation: workloads=NVM dedup workloads; crafted aging workload; baselines=state-of-the-art NVM dedup file systems; metrics=throughput; prefetch gain; metadata read/write amplification; results=1.01–8.98× throughput; metadata R/W amplification 19.35/9.86→6.10/3.43.
Interpretation: 중복 데이터의 공간 지역성을 이용하면 전역 인덱스 접근을 크게 줄일 수 있다.
Reusable lesson: dedup 메타데이터는 데이터 지역성과 같은 단위로 캐시·배치하라.
Applicability: NVM 파일시스템 inline dedup.
Limits: 중복·지역성 정도와 NVM 특성에 민감하며 일부 결과는 crafted aging workload다. +
light dedup a light weight inline deduplication framework for non volatile memory file systems 000ae22e +
Verification: abstract_only; confidence=high.
Canonical title: Light-Dedup: A Light-weight Inline Deduplication Framework for Non-Volatile Memory File Systems
Question: NVM 파일시스템에서 inline dedup의 메타데이터·쓰기 증폭을 줄일 수 있는가?
Context: 전역 fingerprint index는 NVM 접근과 metadata amplification이 크다.
Method: 지역성 인지 region buffer/index와 NVM 내 metadata table, prefetch를 사용한다.
Evaluation: workloads=NVM dedup workloads; crafted aging workload; baselines=state-of-the-art NVM dedup file systems; metrics=throughput; prefetch gain; metadata read/write amplification; results=1.01–8.98× throughput; metadata R/W amplification 19.35/9.86→6.10/3.43.
Interpretation: 중복 데이터의 공간 지역성을 이용하면 전역 인덱스 접근을 크게 줄일 수 있다.
Reusable lesson: dedup 메타데이터는 데이터 지역성과 같은 단위로 캐시·배치하라.
Applicability: NVM 파일시스템 inline dedup.
Limits: 중복·지역성 정도와 NVM 특성에 민감하며 일부 결과는 crafted aging workload다. +
light dedup a light weight inline deduplication framework for non volatile memory file systems 000ae22e +
Bibliographic paper record. +
Bibliographic paper record. +
Verification: latest arXiv abstract/full text, DOI, and official SOSP program; confidence=high.
Canonical title: LithOS: An Operating System for Efficient Machine Learning on GPUs
Question: Can a GPU-native OS isolate and co-schedule ML kernels more efficiently than host-managed sharing?
Context: MPS and existing schedulers have coarse control over thread-block execution, interference, and energy/capacity tradeoffs.
Method: LithOS atomizes kernels, schedules and steals thread-blocks, rightsizes allocations, and manages GPU power.
Evaluation: workloads=stacked inference and hybrid ML workloads; baselines=NVIDIA MPS and prior GPU schedulers; metrics=p99 latency, throughput, capacity, energy; results=13x/3x lower p99 and 1.6x throughput; hybrid 4.7x/1.18x p99 and 1.35x throughput
Interpretation: Thread-block-level OS control exposes isolation and packing opportunities hidden from host schedulers.
Reusable lesson: Schedule accelerators at their native execution unit and jointly manage latency, throughput, and power.
Applicability: Multi-tenant GPUs running inference and mixed ML jobs.
Limits: Requires a specialized GPU OS/runtime and kernel atomization; gains vary by workload interference. +
Verification: latest arXiv abstract/full text, DOI, and official SOSP program; confidence=high.
Canonical title: LithOS: An Operating System for Efficient Machine Learning on GPUs
Question: Can a GPU-native OS isolate and co-schedule ML kernels more efficiently than host-managed sharing?
Context: MPS and existing schedulers have coarse control over thread-block execution, interference, and energy/capacity tradeoffs.
Method: LithOS atomizes kernels, schedules and steals thread-blocks, rightsizes allocations, and manages GPU power.
Evaluation: workloads=stacked inference and hybrid ML workloads; baselines=NVIDIA MPS and prior GPU schedulers; metrics=p99 latency, throughput, capacity, energy; results=13x/3x lower p99 and 1.6x throughput; hybrid 4.7x/1.18x p99 and 1.35x throughput
Interpretation: Thread-block-level OS control exposes isolation and packing opportunities hidden from host schedulers.
Reusable lesson: Schedule accelerators at their native execution unit and jointly manage latency, throughput, and power.
Applicability: Multi-tenant GPUs running inference and mixed ML jobs.
Limits: Requires a specialized GPU OS/runtime and kernel atomization; gains vary by workload interference. +
Verification: official DOI/SOSP metadata and first-author project summary; confidence=medium.
Canonical title: Loom: Efficient Capture and Querying of High-Frequency Telemetry
Question: How can operators capture, correlate, and interactively query complete high-frequency telemetry without pre-aggregation?
Context: Telemetry sources emit at high rates, forcing systems to sample or aggregate away details needed for debugging.
Method: Loom combines an efficient ingest path with lightweight indexes across multiple telemetry sources.
Evaluation: workloads=multiple high-frequency telemetry streams; baselines=; metrics=ingest rate and query response time; results=qualitative high-rate ingest with interactive complete-data queries
Interpretation: A telemetry system can preserve raw correlation value if ingestion and indexing are co-designed.
Reusable lesson: Optimize the write path and defer heavyweight structure while retaining queryable cross-source keys.
Applicability: Performance debugging and observability for high-frequency systems.
Limits: Quantitative workloads, rates, and baselines were not recoverable from accessible primary text. +
Verification: official DOI/SOSP metadata and first-author project summary; confidence=medium.
Canonical title: Loom: Efficient Capture and Querying of High-Frequency Telemetry
Question: How can operators capture, correlate, and interactively query complete high-frequency telemetry without pre-aggregation?
Context: Telemetry sources emit at high rates, forcing systems to sample or aggregate away details needed for debugging.
Method: Loom combines an efficient ingest path with lightweight indexes across multiple telemetry sources.
Evaluation: workloads=multiple high-frequency telemetry streams; baselines=; metrics=ingest rate and query response time; results=qualitative high-rate ingest with interactive complete-data queries
Interpretation: A telemetry system can preserve raw correlation value if ingestion and indexing are co-designed.
Reusable lesson: Optimize the write path and defer heavyweight structure while retaining queryable cross-source keys.
Applicability: Performance debugging and observability for high-frequency systems.
Limits: Quantitative workloads, rates, and baselines were not recoverable from accessible primary text. +
Bibliographic paper record. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_lsm_tree_da534157. 확인 범위: official_abstract. 확인한 자료: https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART002996539 ; https://doi.org/10.5626/KTCP.2023.29.9.444. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_lsm_tree_76287e58. 확인 범위: official_abstract. 확인한 자료: https://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE12318613 ; https://cslab.yonsei.ac.kr/publications. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility. +
m
Bibliographic paper record. +
Verification: official DOI metadata and author project/publication page; confidence=medium.
Canonical title: M5: Mastering Page Migration and Memory Management for CXL-based Tiered Memory Systems
Question: Can the CXL controller identify sparse hot data more precisely than CPU page migration?
Context: CPU schemes can confuse warm with hot pages and waste DRAM on pages containing only a few hot cache lines.
Method: M5 tracks 4 KB page and 64 B word activity in the CXL controller and drives a simple fine-grained migration policy.
Evaluation: workloads=CXL tiered-memory workloads; baselines=best evaluated CPU migration scheme; metrics=hot-data detection and application performance; results=47% more hot data; 14% higher performance
Interpretation: Near-memory visibility supports finer placement decisions without heavy CPU profiling.
Reusable lesson: Track activity at both allocation and transfer granularity when hotness is spatially sparse.
Applicability: CXL DRAM tiers with modifiable controller hardware.
Limits: Requires controller support; gains are measured under a 2–3x CXL/DRAM latency gap. +
Verification: official DOI metadata and author project/publication page; confidence=high.
Canonical title: M5: Hardware-Assisted Fine-Grained Memory Management for CXL-Based Tiered Memory
Question: Can the CXL controller identify sparse hot data more precisely than CPU page migration?
Context: CPU schemes can confuse warm with hot pages and waste DRAM on pages containing only a few hot cache lines.
Method: M5 tracks 4 KB page and 64 B word activity in the CXL controller and drives a simple fine-grained migration policy.
Evaluation: workloads=CXL tiered-memory workloads; baselines=best evaluated CPU migration scheme; metrics=hot-data detection and application performance; results=47% more hot data; 14% higher performance
Interpretation: Near-memory visibility supports finer placement decisions without heavy CPU profiling.
Reusable lesson: Track activity at both allocation and transfer granularity when hotness is spatially sparse.
Applicability: CXL DRAM tiers with modifiable controller hardware.
Limits: Requires controller support; gains are measured under a 2–3x CXL/DRAM latency gap. +