속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
h
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_how_storage_i_o_affects_user_perceived_latency_in_mobile_apps_ac1dc0d2. 확인 범위: metadata_only. 확인한 자료: https://eurosys2015.labri.fr/program/posters/. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Verification: official DOI/SOSP metadata, author page, and institution research report; confidence=medium.
Canonical title: How to Copy Memory? Coordinated Asynchronous Copy as a First-Class OS Service
Question: Can memory copying become a globally optimized asynchronous OS service instead of repeated local memcpy operations?
Context: Copies occur across files, networks, and IPC; synchronous/local handling misses overlap, SIMD/DMA, and redundant-copy elimination.
Method: Copier coordinates copies asynchronously, selects hardware such as SIMD/DMA, and absorbs redundant intermediate copies.
Evaluation: workloads=Redis, proxy, and HarmonyOS-kernel scenarios; baselines=existing copy paths and prior work; metrics=application performance; results=up to 1.8x overall and 1.6x over prior work
Interpretation: Global visibility lets the OS overlap, combine, eliminate, and accelerate copies across subsystem boundaries.
Reusable lesson: Promote ubiquitous data movement to a schedulable service with dependency-aware completion.
Applicability: Copy-heavy file, network, IPC, and mobile OS paths.
Limits: Applications/subsystems must express asynchronous copy dependencies; hardware availability affects benefit. +
Bibliographic paper record. +
Verification: official DOI/SOSP metadata, author page, and institution research report; confidence=high.
Canonical title: How to Copy Memory? Coordinated Asynchronous Copy as a First-Class OS Service
Question: Can memory copying become a globally optimized asynchronous OS service instead of repeated local memcpy operations?
Context: Copies occur across files, networks, and IPC; synchronous/local handling misses overlap, SIMD/DMA, and redundant-copy elimination.
Method: Copier coordinates copies asynchronously, selects hardware such as SIMD/DMA, and absorbs redundant intermediate copies.
Evaluation: workloads=Redis, proxy, and HarmonyOS-kernel scenarios; baselines=existing copy paths and prior work; metrics=application performance; results=up to 1.8x overall and 1.6x over prior work
Interpretation: Global visibility lets the OS overlap, combine, eliminate, and accelerate copies across subsystem boundaries.
Reusable lesson: Promote ubiquitous data movement to a schedulable service with dependency-aware completion.
Applicability: Copy-heavy file, network, IPC, and mobile OS paths.
Limits: Applications/subsystems must express asynchronous copy dependencies; hardware availability affects benefit. +
i
Bibliographic paper record. +
Verification: abstract_only; confidence=medium.
Canonical title: I/O Passthru: Upstreaming a Flexible and Efficient I/O Path in Linux
Question: Linux 범용 블록 계층 비용을 우회하는 유연한 NVMe 경로를 mainline에 넣을 수 있는가?
Context: 고성능 애플리케이션은 범용 블록 경로의 변환·락·큐잉 비용을 부담한다.
Method: NVMe character device와 io_uring_cmd로 사용자 명령을 장치에 직접 전달한다.
Evaluation: workloads=FIO; CacheLib data placement; baselines=Linux block I/O path; metrics=peak IOPS; results=16–40% higher peak IOPS.
Interpretation: 표준 비동기 API 안에서 선택적 bypass를 제공하면 성능과 upstream 유지보수를 함께 얻는다.
Reusable lesson: fast path는 별도 생태계보다 기존 커널 API에 통합하라.
Applicability: NVMe 최적화 DB·캐시·스토리지 엔진.
Limits: 범용 블록 서비스 일부를 우회하고 NVMe·peak FIO 중심 결과다. +
Verification: abstract_only; confidence=high.
Canonical title: I/O Passthru: Upstreaming a Flexible and Efficient I/O Path in Linux
Question: Linux 범용 블록 계층 비용을 우회하는 유연한 NVMe 경로를 mainline에 넣을 수 있는가?
Context: 고성능 애플리케이션은 범용 블록 경로의 변환·락·큐잉 비용을 부담한다.
Method: NVMe character device와 io_uring_cmd로 사용자 명령을 장치에 직접 전달한다.
Evaluation: workloads=FIO; CacheLib data placement; baselines=Linux block I/O path; metrics=peak IOPS; results=16–40% higher peak IOPS.
Interpretation: 표준 비동기 API 안에서 선택적 bypass를 제공하면 성능과 upstream 유지보수를 함께 얻는다.
Reusable lesson: fast path는 별도 생태계보다 기존 커널 API에 통합하라.
Applicability: NVMe 최적화 DB·캐시·스토리지 엔진.
Limits: 범용 블록 서비스 일부를 우회하고 NVMe·peak FIO 중심 결과다. +
Verification: arXiv abstract/full text and official DOI metadata; confidence=high.
Canonical title: IC-Cache: Efficient Large Language Model Serving via In-context Caching
Question: Can prior question-answer results be reused as in-context examples to reduce expensive large-model serving?
Context: Many incoming queries resemble historical requests, but conventional caches require exact matches.
Method: IC-Cache selects historical Q/A examples, routes requests adaptively, and replays them with a cost-aware policy.
Evaluation: workloads=millions of realistic LLM requests; baselines=uncached and conventional serving; metrics=similarity coverage, throughput, latency, quality; results=>70% reusable counterparts; 1.4–5.9x throughput; 28–71% lower latency; no quality loss
Interpretation: Semantic history can be an executable cache that substitutes cheaper conditioned inference for repeated full reasoning.
Reusable lesson: Cache reusable demonstrations, then route by similarity, quality risk, and serving cost.
Applicability: High-volume LLM services with recurrent intents or tasks.
Limits: Requires representative history and safe example selection; novel or distribution-shifted queries may not benefit. +
Bibliographic paper record. +
Verification: arXiv abstract/full text and official DOI metadata; confidence=high.
Canonical title: IC-Cache: Efficient Large Language Model Serving via In-context Caching
Question: Can prior question-answer results be reused as in-context examples to reduce expensive large-model serving?
Context: Many incoming queries resemble historical requests, but conventional caches require exact matches.
Method: IC-Cache selects historical Q/A examples, routes requests adaptively, and replays them with a cost-aware policy.
Evaluation: workloads=millions of realistic LLM requests; baselines=uncached and conventional serving; metrics=similarity coverage, throughput, latency, quality; results=>70% reusable counterparts; 1.4–5.9x throughput; 28–71% lower latency; no quality loss
Interpretation: Semantic history can be an executable cache that substitutes cheaper conditioned inference for repeated full reasoning.
Reusable lesson: Cache reusable demonstrations, then route by similarity, quality risk, and serving cost.
Applicability: High-volume LLM services with recurrent intents or tasks.
Limits: Requires representative history and safe example selection; novel or distribution-shifted queries may not benefit. +
ice collaborating memory and process management for user experience on resource limited mobile d 312e453f +
Bibliographic paper record. +
ice collaborating memory and process management for user experience on resource limited mobile d 312e453f +
Verification: conference_metadata_plus_primary_followup_abstract; confidence=medium.
Canonical title: ICE: Collaborating Memory and Process Management for User Experience on Resource-limited Mobile Devices
Question: Can mobile memory and process management cooperate to stop background refault activity from disrupting foreground frame rendering?
Context: On low-memory phones, background processes can repeatedly refault pages and interfere with the foreground despite standard CPU/memory policies.
Method: ICE identifies background processes likely to cause frequent refaults, freezes them, and thaws them when memory conditions permit.
Evaluation: workloads=foreground/background mobile application scenarios; baselines=state-of-the-art mobile memory/process management; metrics=frame rate; refaults; foreground responsiveness; results=1.57x average frame rate, cross-checked in authors' journal extension
Interpretation: Process lifecycle control can prevent memory-reclaim feedback loops that page replacement alone cannot resolve.
Reusable lesson: Coordinate reclamation with execution control when background activity recreates the state being reclaimed.
Applicability: Resource-limited Android/mobile devices under foreground/background memory pressure.
Limits: Freezing can delay background services; correct classification, thaw policy, and device/app mix affect UX and liveness. +
ice collaborating memory and process management for user experience on resource limited mobile d 312e453f +
Verification: conference_metadata_plus_primary_followup_abstract; confidence=medium.
Canonical title: ICE: Collaborating Memory and Process Management for User Experience on Resource-limited Mobile Devices
Question: Can mobile memory and process management cooperate to stop background refault activity from disrupting foreground frame rendering?
Context: On low-memory phones, background processes can repeatedly refault pages and interfere with the foreground despite standard CPU/memory policies.
Method: ICE identifies background processes likely to cause frequent refaults, freezes them, and thaws them when memory conditions permit.
Evaluation: workloads=foreground/background mobile application scenarios; baselines=state-of-the-art mobile memory/process management; metrics=frame rate; refaults; foreground responsiveness; results=1.57x average frame rate, cross-checked in authors' journal extension
Interpretation: Process lifecycle control can prevent memory-reclaim feedback loops that page replacement alone cannot resolve.
Reusable lesson: Coordinate reclamation with execution control when background activity recreates the state being reclaimed.
Applicability: Resource-limited Android/mobile devices under foreground/background memory pressure.
Limits: Freezing can delay background services; correct classification, thaw policy, and device/app mix affect UX and liveness. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_identifying_on_off_cpu_bottlenecks_together_with_blocked_samples_2f1283b1. 확인 범위: full_text. 확인한 자료: https://www.usenix.org/conference/osdi24/presentation/ahn ; https://www.usenix.org/system/files/osdi24-ahn.pdf. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1. +
impress an importance informed multi tier prefix kv storage system for large language model infe 51b47015 +
Verification: official_abstract; confidence=high.
Canonical title: IMPRESS: An Importance-Informed Multi-Tier Prefix KV Storage System for Large Language Model Inference
Question: How can disk-tiered prefix KV reuse reduce LLM time to first token when loading every cached token is too slow?
Context: Repeated long prefixes save prefill compute, but CPU-memory limits push KV state to disks whose I/O can erase the benefit.
Method: IMPRESS exploits similarity in important-token indices across attention heads to identify and load only important KVs, then manages prefix state across storage/cache tiers by importance.
Evaluation: workloads=LLM inference with reusable long prefixes; baselines=state-of-the-art prefix KV storage systems; metrics=TTFT; inference accuracy; KV I/O; results=up to 2.8x lower TTFT with comparable accuracy
Interpretation: Approximate semantic importance can reduce KV I/O more effectively than indiscriminate prefix restoration.
Reusable lesson: When cached state is too large to reload, rank units by output importance and tier them accordingly.
Applicability: LLM applications with recurring long contexts and disk-backed prefix caches.
Limits: Accuracy and speed depend on importance stability across heads/models/tasks, prefix reuse, disk bandwidth, and tier capacity. +
impress an importance informed multi tier prefix kv storage system for large language model infe 51b47015 +
Bibliographic paper record. +
impress an importance informed multi tier prefix kv storage system for large language model infe 51b47015 +
Verification: official_abstract; confidence=medium.
Canonical title: IMPRESS: An Importance-Informed Multi-Tier Prefix KV Storage System for Large Language Model Inference
Question: How can disk-tiered prefix KV reuse reduce LLM time to first token when loading every cached token is too slow?
Context: Repeated long prefixes save prefill compute, but CPU-memory limits push KV state to disks whose I/O can erase the benefit.
Method: IMPRESS exploits similarity in important-token indices across attention heads to identify and load only important KVs, then manages prefix state across storage/cache tiers by importance.
Evaluation: workloads=LLM inference with reusable long prefixes; baselines=state-of-the-art prefix KV storage systems; metrics=TTFT; inference accuracy; KV I/O; results=up to 2.8x lower TTFT with comparable accuracy
Interpretation: Approximate semantic importance can reduce KV I/O more effectively than indiscriminate prefix restoration.
Reusable lesson: When cached state is too large to reload, rank units by output importance and tier them accordingly.
Applicability: LLM applications with recurring long contexts and disk-backed prefix caches.
Limits: Accuracy and speed depend on importance stability across heads/models/tasks, prefix reuse, disk bandwidth, and tier capacity. +
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_improving_application_launch_performance_in_smartphones_using_recurrent_neural_47c597b7. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1145/3231884.3231897 ; https://yonsei.elsevierpure.com/en/publications/improving-application-launch-performance-in-smartphones-using-rec-2/. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +