속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
b
Bibliographic paper record. +
Verification: official DOI/ASPLOS abstract and arXiv paper; confidence=high.
Canonical title: BypassD: Enabling Fast Userspace Access to Shared SSDs
Question: Can unmodified applications access a shared SSD from userspace at near-SPDK latency while retaining protection and files?
Context: Kernel I/O is safe/shareable but slow; kernel-bypass frameworks require exclusive devices or application rewrites.
Method: BypassD exposes virtual-address access and uses the IOMMU to translate file offsets to LBAs on a shared SSD.
Evaluation: workloads=4 KB I/O microbenchmarks and WiredTiger; baselines=Linux and SPDK; metrics=latency and application performance; results=42% lower 4 KB latency vs Linux; near SPDK; about 20% WiredTiger improvement
Interpretation: The IOMMU can enforce storage address translation/protection for a shared userspace fast path.
Reusable lesson: Move per-I/O translation into hardware while retaining kernel control-plane setup.
Applicability: Shared NVMe SSDs serving unmodified storage applications.
Limits: Depends on suitable IOMMU/device support and a setup/control path in the OS. +
Verification: official DOI/ASPLOS abstract and arXiv paper; confidence=high.
Canonical title: BypassD: Enabling Fast Userspace Access to Shared SSDs
Question: Can unmodified applications access a shared SSD from userspace at near-SPDK latency while retaining protection and files?
Context: Kernel I/O is safe/shareable but slow; kernel-bypass frameworks require exclusive devices or application rewrites.
Method: BypassD exposes virtual-address access and uses the IOMMU to translate file offsets to LBAs on a shared SSD.
Evaluation: workloads=4 KB I/O microbenchmarks and WiredTiger; baselines=Linux and SPDK; metrics=latency and application performance; results=42% lower 4 KB latency vs Linux; near SPDK; about 20% WiredTiger improvement
Interpretation: The IOMMU can enforce storage address translation/protection for a shared userspace fast path.
Reusable lesson: Move per-I/O translation into hardware while retaining kernel control-plane setup.
Applicability: Shared NVMe SSDs serving unmodified storage applications.
Limits: Depends on suitable IOMMU/device support and a setup/control path in the OS. +
Bibliographic paper record. +
byteexpress a high performance and traffic efficient inline transfer of small payloads over nvme a7c01a49 +
Verification: official DOI metadata and author publication page/abstract; confidence=medium.
Canonical title: ByteExpress: A High-Performance and Traffic-Efficient Inline Transfer of Small Payloads over NVMe
Question: Can sub-page NVMe payloads avoid separate DMA data transfers?
Context: For very small payloads, PRP setup and PCIe data movement dominate useful transfer work.
Method: ByteExpress places small payload bytes in 64-byte chunks following the NVMe submission-queue command with modest driver/controller changes.
Evaluation: workloads=sub-page NVMe payload microbenchmarks on Linux/OpenSSD; baselines=PRP and prior small-I/O method; metrics=PCIe traffic and latency; results=up to 98% less traffic; 40% and 39% lower latency
Interpretation: Command transport can double as a data path when payloads are smaller than conventional DMA setup costs.
Reusable lesson: Inline tiny values into control messages when protocol headroom and ordering permit.
Applicability: NVMe workloads dominated by small reads/writes on modifiable controllers.
Limits: Prototype uses OpenSSD and targets sub-page payloads; benefit does not generalize to large transfers. +
byteexpress a high performance and traffic efficient inline transfer of small payloads over nvme a7c01a49 +
Verification: official DOI metadata and author publication page/abstract; confidence=high.
Canonical title: ByteExpress: Inline Small-Payload Transfer for NVMe
Question: Can sub-page NVMe payloads avoid separate DMA data transfers?
Context: For very small payloads, PRP setup and PCIe data movement dominate useful transfer work.
Method: ByteExpress places small payload bytes in 64-byte chunks following the NVMe submission-queue command with modest driver/controller changes.
Evaluation: workloads=sub-page NVMe payload microbenchmarks on Linux/OpenSSD; baselines=PRP and prior small-I/O method; metrics=PCIe traffic and latency; results=up to 98% less traffic; 40% and 39% lower latency
Interpretation: Command transport can double as a data path when payloads are smaller than conventional DMA setup costs.
Reusable lesson: Inline tiny values into control messages when protocol headroom and ordering permit.
Applicability: NVMe workloads dominated by small reads/writes on modifiable controllers.
Limits: Prototype uses OpenSSD and targets sub-page payloads; benefit does not generalize to large transfers. +
byteexpress a high performance and traffic efficient inline transfer of small payloads over nvme a7c01a49 +
Bibliographic paper record. +
c
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_cache_scheme_of_shared_buffer_mappings_for_energy_efficiency_of_mobile_devices_f1dee859. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1049/el.2015.0244 ; https://yonsei.elsevierpure.com/en/publications/cache-scheme-of-shared-buffer-mappings-for-energy-efficiency-of-m/. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Verification: arXiv abstract/full-text excerpts and DOI metadata; confidence=medium.
Canonical title: CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion
Question: Can RAG reuse cached KV states for retrieved chunks without losing cross-chunk attention quality?
Context: Full recomputation is slow, while naive KV reuse omits interactions created when chunks are concatenated.
Method: CacheBlend selectively recomputes fewer than 15% of tokens in typical cases and pipelines recomputation with cache fetch.
Evaluation: workloads=three open-source LLMs; four QA/RAG/summarization datasets; baselines=full KV recomputation and naive full reuse; metrics=quality, TTFT, throughput, recomputed-token share; results=<15% typical recomputation; 2.2–3.3x TTFT and 2.8–5x throughput improvement
Interpretation: Most cached chunk states remain usable if a small, well-chosen subset repairs cross-context interactions.
Reusable lesson: Use selective correction between exact recomputation and unsafe cache reuse.
Applicability: RAG services with recurring document chunks and reusable KV caches.
Limits: Requires cache storage and token selection; benefits depend on chunk reuse and retrieval composition. +
Verification: arXiv abstract/full-text excerpts and DOI metadata; confidence=high.
Canonical title: CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion
Question: Can RAG reuse cached KV states for retrieved chunks without losing cross-chunk attention quality?
Context: Full recomputation is slow, while naive KV reuse omits interactions created when chunks are concatenated.
Method: CacheBlend selectively recomputes fewer than 15% of tokens in typical cases and pipelines recomputation with cache fetch.
Evaluation: workloads=three open-source LLMs; four QA/RAG/summarization datasets; baselines=full KV recomputation and naive full reuse; metrics=quality, TTFT, throughput, recomputed-token share; results=<15% typical recomputation; 2.2–3.3x TTFT and 2.8–5x throughput improvement
Interpretation: Most cached chunk states remain usable if a small, well-chosen subset repairs cross-context interactions.
Reusable lesson: Use selective correction between exact recomputation and unsafe cache reuse.
Applicability: RAG services with recurring document chunks and reusable KV caches.
Limits: Requires cache storage and token selection; benefits depend on chunk reuse and retrieval composition. +
Bibliographic paper record. +
Verification: abstract_only; confidence=high.
Canonical title: Calcspar: A Contract-Aware LSM Store for Cloud Storage with Low Latency Spikes
Question: EBS의 계약·지연 변동 속에서도 LSM 읽기 꼬리 지연을 제한할 수 있는가?
Context: I/O 압력과 LSM 내부 작업이 클라우드 블록 저장소의 지연 스파이크를 증폭한다.
Method: 요청률 조절, 변동 인지 캐시, 혼잡 인지 IOPS 할당기로 외부·내부 I/O를 제어한다.
Evaluation: workloads=Amazon EBS; real-world workloads; baselines=state-of-the-art LSM stores; metrics=p99 latency; average latency; read/write performance; results=p99 <550 μs; average latency -66%.
Interpretation: 스토리지 계약과 순간 혼잡을 명시적으로 모델링하면 LSM 꼬리 지연을 안정화할 수 있다.
Reusable lesson: 클라우드 스토리지 최적화에는 계약 한도와 변동성 모두를 반영하라.
Applicability: EBS 위 LSM KV/DB.
Limits: EBS와 선정 워크로드 중심이며 캐시·IOPS 정책의 이식성은 미검증이다. +
Verification: abstract_only; confidence=medium.
Canonical title: Calcspar: A Contract-Aware LSM Store for Cloud Storage with Low Latency Spikes
Question: EBS의 계약·지연 변동 속에서도 LSM 읽기 꼬리 지연을 제한할 수 있는가?
Context: I/O 압력과 LSM 내부 작업이 클라우드 블록 저장소의 지연 스파이크를 증폭한다.
Method: 요청률 조절, 변동 인지 캐시, 혼잡 인지 IOPS 할당기로 외부·내부 I/O를 제어한다.
Evaluation: workloads=Amazon EBS; real-world workloads; baselines=state-of-the-art LSM stores; metrics=p99 latency; average latency; read/write performance; results=p99 <550 μs; average latency -66%.
Interpretation: 스토리지 계약과 순간 혼잡을 명시적으로 모델링하면 LSM 꼬리 지연을 안정화할 수 있다.
Reusable lesson: 클라우드 스토리지 최적화에는 계약 한도와 변동성 모두를 반영하라.
Applicability: EBS 위 LSM KV/DB.
Limits: EBS와 선정 워크로드 중심이며 캐시·IOPS 정책의 이식성은 미검증이다. +
Bibliographic paper record. +
Bibliographic paper record. +
Verification: official_abstract; confidence=high.
Canonical title: Can ZNS SSDs be Better Storage Devices for Persistent Cache?
Question: Which integration design lets persistent caches exploit ZNS SSD capacity and host-controlled garbage collection despite update-heavy writes?
Context: Block SSD cache updates trigger internal garbage collection and write amplification; ZNS exposes sequential zones but requires host management.
Method: The study compares three adaptations: a zone-aware file-system cache, a cache-unit-to-zone aligned design, and a lightweight region-management layer between CacheLib and ZNS.
Evaluation: workloads=Cachebench; RocksDB with ZNS secondary cache; baselines=compatible regular SSD; F2FS-based cache; three ZNS integration schemes; metrics=throughput; cache hit ratio; write amplification; capacity; results=qualitative tradeoff study; no numeric result in primary abstract
Interpretation: ZNS benefits are largest when cache eviction/allocation and zone lifecycle are co-designed rather than hidden below a block interface.
Reusable lesson: Align application reclamation units with device erase/management units when the host controls placement.
Applicability: CacheLib-like persistent flash caches on ZNS SSDs.
Limits: Workshop exploration with three schemes; results depend on cache-unit/zone sizes, workload locality, ZNS device, and recovery design. +
Verification: official_abstract; confidence=medium.
Canonical title: Can ZNS SSDs be Better Storage Devices for Persistent Cache?
Question: Which integration design lets persistent caches exploit ZNS SSD capacity and host-controlled garbage collection despite update-heavy writes?
Context: Block SSD cache updates trigger internal garbage collection and write amplification; ZNS exposes sequential zones but requires host management.
Method: The study compares three adaptations: a zone-aware file-system cache, a cache-unit-to-zone aligned design, and a lightweight region-management layer between CacheLib and ZNS.
Evaluation: workloads=Cachebench; RocksDB with ZNS secondary cache; baselines=compatible regular SSD; F2FS-based cache; three ZNS integration schemes; metrics=throughput; cache hit ratio; write amplification; capacity; results=qualitative tradeoff study; no numeric result in primary abstract
Interpretation: ZNS benefits are largest when cache eviction/allocation and zone lifecycle are co-designed rather than hidden below a block interface.
Reusable lesson: Align application reclamation units with device erase/management units when the host controls placement.
Applicability: CacheLib-like persistent flash caches on ZNS SSDs.
Limits: Workshop exploration with three schemes; results depend on cache-unit/zone sizes, workload locality, ZNS device, and recovery design. +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_catching_two_rabbits_adaptive_real_time_support_for_embedded_linux_07b12680. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1002/spe.911 ; https://yonsei.elsevierpure.com/en/publications/catching-two-rabbits-adaptive-real-time-support-for-embedded-linu/ ; https://koasas.kaist.ac.kr/handle/10203/93501. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +