속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
k
Bibliographic paper record. +
Verification: abstract_only; confidence=medium.
Canonical title: Kosmo: Efficient Online Miss Ratio Curve Generation for Eviction Policy Evaluation
Question: 비-inclusion 축출 정책의 miss-ratio curve를 온라인·저비용으로 만들 수 있는가?
Context: 여러 캐시 크기를 따로 시뮬레이션하면 메모리·연산 비용이 크다.
Method: Kosmo는 정책별 상태를 공유·근사해 여러 용량의 MRC를 동시에 생성한다.
Evaluation: workloads=52 public traces; ~126 billion accesses; baselines=Miniature Simulations with 100 caches; metrics=memory; throughput; MRC accuracy; results=3.6× average/36× max less memory; 1.3× throughput.
Interpretation: 정책 평가도 전체 시뮬레이션 대신 공유 가능한 충분 통계로 축약할 수 있다.
Reusable lesson: 캐시 용량 탐색은 여러 독립 시뮬레이터보다 상태 공유형 근사를 사용하라.
Applicability: 온라인 캐시 크기·정책 튜닝.
Limits: 근사 오차와 샘플링, 지원 정책 집합에 제약된다. +
Verification: abstract_only; confidence=high.
Canonical title: Kosmo: Efficient Online Miss Ratio Curve Generation for Eviction Policy Evaluation
Question: 비-inclusion 축출 정책의 miss-ratio curve를 온라인·저비용으로 만들 수 있는가?
Context: 여러 캐시 크기를 따로 시뮬레이션하면 메모리·연산 비용이 크다.
Method: Kosmo는 정책별 상태를 공유·근사해 여러 용량의 MRC를 동시에 생성한다.
Evaluation: workloads=52 public traces; ~126 billion accesses; baselines=Miniature Simulations with 100 caches; metrics=memory; throughput; MRC accuracy; results=3.6× average/36× max less memory; 1.3× throughput.
Interpretation: 정책 평가도 전체 시뮬레이션 대신 공유 가능한 충분 통계로 축약할 수 있다.
Reusable lesson: 캐시 용량 탐색은 여러 독립 시뮬레이터보다 상태 공유형 근사를 사용하라.
Applicability: 온라인 캐시 크기·정책 튜닝.
Limits: 근사 오차와 샘플링, 지원 정책 집합에 제약된다. +
kvaccel a novel write accelerator for lsm tree based kv stores with host ssd collaboration 968f5966 +
Verification: conference paper PDF, DOI metadata, and arXiv preprint; confidence=high.
Canonical title: KVACCEL: A Novel Write Accelerator for LSM-Tree-Based KV Stores with Host-SSD Collaboration
Question: Can one SSD support both block and key-value paths while accelerating LSM-tree writes without violating consistency?
Context: Host-managed LSM stores suffer write stalls and overhead, while KV SSDs alone do not preserve the conventional block interface.
Method: KVACCEL partitions NAND between block and KV interfaces, uses the KV region as a temporary write buffer, and adds an in-device LSM buffer plus range-scan iterator.
Evaluation: workloads=write-intensive and mixed read/write key-value workloads; baselines=ADOC and conventional paths; metrics=throughput and CPU efficiency; results=up to 17% throughput gain on write-intensive workloads; comparable mixed-workload performance
Interpretation: A KV namespace can act as a pressure-relief tier for the host LSM while block compatibility is retained.
Reusable lesson: Use device interfaces according to data lifetime and write-path urgency, with an explicit reconciliation protocol.
Applicability: LSM-based stores on programmable dual-interface SSDs.
Limits: Requires modified dual-interface SSD firmware; the headline gain is for write-intensive workloads. +
kvaccel a novel write accelerator for lsm tree based kv stores with host ssd collaboration 968f5966 +
Bibliographic paper record. +
kvaccel a novel write accelerator for lsm tree based kv stores with host ssd collaboration 968f5966 +
Verification: conference paper PDF, DOI metadata, and arXiv preprint; confidence=high.
Canonical title: KVACCEL: A Dual-Interface SSD Architecture for Key-Value Storage Acceleration
Question: Can one SSD support both block and key-value paths while accelerating LSM-tree writes without violating consistency?
Context: Host-managed LSM stores suffer write stalls and overhead, while KV SSDs alone do not preserve the conventional block interface.
Method: KVACCEL partitions NAND between block and KV interfaces, uses the KV region as a temporary write buffer, and adds an in-device LSM buffer plus range-scan iterator.
Evaluation: workloads=write-intensive and mixed read/write key-value workloads; baselines=ADOC and conventional paths; metrics=throughput and CPU efficiency; results=up to 17% throughput gain on write-intensive workloads; comparable mixed-workload performance
Interpretation: A KV namespace can act as a pressure-relief tier for the host LSM while block compatibility is retained.
Reusable lesson: Use device interfaces according to data lifetime and write-path urgency, with an explicit reconciliation protocol.
Applicability: LSM-based stores on programmable dual-interface SSDs.
Limits: Requires modified dual-interface SSD firmware; the headline gain is for write-intensive workloads. +
kvcache cache in the wild characterizing and optimizing kvcache cache at a large cloud provider 0b39b53c +
Verification: official USENIX paper PDF and arXiv abstract; confidence=high.
Canonical title: KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider
Question: What reuse patterns actually determine KV-cache value in production LLM traffic?
Context: KV-cache policies are often designed from synthetic multi-turn assumptions rather than cloud traces.
Method: The paper analyzes large-provider KV-cache traces and derives category-aware, workload-aware caching and eviction guidance.
Evaluation: workloads=large cloud-provider KV-cache traces; baselines=generic cache sizing and eviction assumptions; metrics=reuse incidence, predictability, and ideal capacity; results=qualitative production reuse findings; no unambiguous primary headline number recorded
Interpretation: Production cache value depends on semantic/request categories, not conversation length alone.
Reusable lesson: Measure real reuse classes before choosing cache size or eviction policy.
Applicability: Shared LLM serving clusters with cross-request prefix/KV reuse.
Limits: Trace findings may reflect one provider and period; exact intervention gains were not extracted from the primary PDF. +
kvcache cache in the wild characterizing and optimizing kvcache cache at a large cloud provider 0b39b53c +
Bibliographic paper record. +
kvcache cache in the wild characterizing and optimizing kvcache cache at a large cloud provider 0b39b53c +
Verification: official USENIX paper PDF and arXiv abstract; confidence=high.
Canonical title: KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider
Question: What reuse patterns actually determine KV-cache value in production LLM traffic?
Context: KV-cache policies are often designed from synthetic multi-turn assumptions rather than cloud traces.
Method: The paper analyzes large-provider KV-cache traces and derives category-aware, workload-aware caching and eviction guidance.
Evaluation: workloads=large cloud-provider KV-cache traces; baselines=generic cache sizing and eviction assumptions; metrics=reuse incidence, predictability, and ideal capacity; results=qualitative production reuse findings; no unambiguous primary headline number recorded
Interpretation: Production cache value depends on semantic/request categories, not conversation length alone.
Reusable lesson: Measure real reuse classes before choosing cache size or eviction policy.
Applicability: Shared LLM serving clusters with cross-request prefix/KV reuse.
Limits: Trace findings may reflect one provider and period; exact intervention gains were not extracted from the primary PDF. +
l
Verification: official USENIX page and abstract; confidence=high.
Canonical title: LegoOS: A Disseminated, Distributed OS for Hardware Resource Disaggregation
Question: What OS architecture can present a coherent machine abstraction over disaggregated processors, memory, and storage?
Context: Monolithic kernels assume resources share one server and do not naturally support independent scaling or failure domains.
Method: LegoOS uses a splitkernel with monitors colocated with each disaggregated resource and assembles virtual nodes for applications.
Evaluation: workloads=applications on an emulated disaggregated cluster; baselines=Linux; metrics=performance, packing, failure rate; results=Linux-comparable performance; qualitative packing/failure improvements
Interpretation: Disaggregation benefits from distributing OS responsibilities to the resource that owns the relevant state.
Reusable lesson: Split kernel services along hardware ownership boundaries while preserving a familiar application abstraction.
Applicability: Rack-scale disaggregated compute, memory, and storage fabrics.
Limits: Evaluation used an emulated commodity-server environment rather than native disaggregated hardware. +
Bibliographic paper record. +
Verification: official USENIX page and abstract; confidence=medium.
Canonical title: LegoOS: A Disseminated, Distributed OS for Hardware Resource Disaggregation
Question: What OS architecture can present a coherent machine abstraction over disaggregated processors, memory, and storage?
Context: Monolithic kernels assume resources share one server and do not naturally support independent scaling or failure domains.
Method: LegoOS uses a splitkernel with monitors colocated with each disaggregated resource and assembles virtual nodes for applications.
Evaluation: workloads=applications on an emulated disaggregated cluster; baselines=Linux; metrics=performance, packing, failure rate; results=Linux-comparable performance; qualitative packing/failure improvements
Interpretation: Disaggregation benefits from distributing OS responsibilities to the resource that owns the relevant state.
Reusable lesson: Split kernel services along hardware ownership boundaries while preserving a familiar application abstraction.
Applicability: Rack-scale disaggregated compute, memory, and storage fabrics.
Limits: Evaluation used an emulated commodity-server environment rather than native disaggregated hardware. +
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_56ee25b0. 확인 범위: official_abstract. 확인한 자료: https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART001685500 ; 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_5b128265. 확인 범위: official_abstract. 확인한 자료: https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART002164736 ; 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_bdff2f7c. 확인 범위: official_abstract. 확인한 자료: https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART001899004 ; https://doi.org/10.3745/KTCCS.2014.3.7.209. 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_fd6afd43. 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. +