속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
c
Verification: full_text; confidence=high.
Canonical title: Cornus: Atomic Commit for a Cloud DBMS with Storage Disaggregation
Question: 분리형 저장소 기반 클라우드 DBMS의 원자적 커밋 지연과 장애 복구를 줄일 수 있는가?
Context: 전통적 2PC의 코디네이터 결정 로그가 원격 저장소 왕복과 단일 병목을 만든다.
Method: Cornus는 저장소의 원자적 LogOnce(CAS) 연산으로 참가자 로그를 조정해 코디네이터 결정 로그를 없앤다.
Evaluation: workloads=Sundial on Azure Blob; Sundial on Redis; baselines=classic 2PC; metrics=commit latency; throughput; results=Up to 1.9× latency speedup.
Interpretation: 고가용 분리형 저장소를 합의 보조자로 활용하면 커밋 경로를 단축할 수 있다.
Reusable lesson: 저장소가 제공하는 원자 연산을 프로토콜 설계에 직접 반영하라.
Applicability: 스토리지 분리형 트랜잭션 DB, 클라우드 OLTP.
Limits: 고가용 저장소와 LogOnce/CAS 의미론에 의존하며 저장소 지연에 민감하다. +
Bibliographic paper record. +
Verification: full_text; confidence=high.
Canonical title: Cornus: Atomic Commit for a Cloud DBMS with Storage Disaggregation
Question: 분리형 저장소 기반 클라우드 DBMS의 원자적 커밋 지연과 장애 복구를 줄일 수 있는가?
Context: 전통적 2PC의 코디네이터 결정 로그가 원격 저장소 왕복과 단일 병목을 만든다.
Method: Cornus는 저장소의 원자적 LogOnce(CAS) 연산으로 참가자 로그를 조정해 코디네이터 결정 로그를 없앤다.
Evaluation: workloads=Sundial on Azure Blob; Sundial on Redis; baselines=classic 2PC; metrics=commit latency; throughput; results=Up to 1.9× latency speedup.
Interpretation: 고가용 분리형 저장소를 합의 보조자로 활용하면 커밋 경로를 단축할 수 있다.
Reusable lesson: 저장소가 제공하는 원자 연산을 프로토콜 설계에 직접 반영하라.
Applicability: 스토리지 분리형 트랜잭션 DB, 클라우드 OLTP.
Limits: 고가용 저장소와 LogOnce/CAS 의미론에 의존하며 저장소 지연에 민감하다. +
Verification: full_text; confidence=high.
Question: Can main-memory OLTP hide pointer-chasing stalls without rewriting the database around manual continuations?
Context: Modern in-memory engines are often memory-latency bound, but conventional batching complicates transaction code.
Method: CoroBase expresses each transaction as a coroutine and interleaves transactions to batch memory accesses and issue software prefetches while retaining the transaction API.
Evaluation: workloads=read-intensive and mixed OLTP on a 48-core server; baselines=conventional main-memory execution; metrics=transaction throughput; scalability; results=close to 2x on read-intensive workloads
Interpretation: Language/runtime continuations can hide latency while preserving readable transactional control flow.
Reusable lesson: Use coroutines to interleave independent requests around predictable long-latency accesses.
Applicability: Main-memory databases with pointer-heavy read paths.
Limits: Benefits shrink for compute/write-heavy workloads and depend on useful prefetch distance and batching concurrency. +
Bibliographic paper record. +
Verification: full_text; confidence=high.
Question: Can main-memory OLTP hide pointer-chasing stalls without rewriting the database around manual continuations?
Context: Modern in-memory engines are often memory-latency bound, but conventional batching complicates transaction code.
Method: CoroBase expresses each transaction as a coroutine and interleaves transactions to batch memory accesses and issue software prefetches while retaining the transaction API.
Evaluation: workloads=read-intensive and mixed OLTP on a 48-core server; baselines=conventional main-memory execution; metrics=transaction throughput; scalability; results=close to 2x on read-intensive workloads
Interpretation: Language/runtime continuations can hide latency while preserving readable transactional control flow.
Reusable lesson: Use coroutines to interleave independent requests around predictable long-latency accesses.
Applicability: Main-memory databases with pointer-heavy read paths.
Limits: Benefits shrink for compute/write-heavy workloads and depend on useful prefetch distance and batching concurrency. +
cost efficient large language model serving for multi turn conversations with cachedattention aa200b5c +
Verification: official_abstract; confidence=medium.
Question: How can multi-turn LLM services reuse prior-turn KV state without consuming scarce GPU memory?
Context: Recomputing the conversation prefix wastes prefill work; retaining every KV cache on GPU limits concurrency.
Method: CachedAttention stores KV state hierarchically, preloads and saves it layer by layer, coordinates fetch/eviction with scheduling, and decouples position encoding from truncation.
Evaluation: workloads=multi-turn LLM conversations; baselines=recompute and GPU-resident KV-cache serving; metrics=TTFT; prefill throughput; end-to-end cost; results=up to 87% lower TTFT; up to 7.8x prefill throughput; up to 70% lower cost
Interpretation: KV persistence is effective when storage movement is overlapped at layer granularity and made scheduler-visible.
Reusable lesson: Persist reusable intermediate state below the accelerator and pipeline its movement with computation.
Applicability: Multi-turn chatbot and agent services with repeated conversation prefixes.
Limits: Benefits depend on prefix reuse, storage bandwidth/capacity, context truncation, and scheduler locality. +
cost efficient large language model serving for multi turn conversations with cachedattention aa200b5c +
Bibliographic paper record. +
cost efficient large language model serving for multi turn conversations with cachedattention aa200b5c +
Verification: official_abstract; confidence=high.
Question: How can multi-turn LLM services reuse prior-turn KV state without consuming scarce GPU memory?
Context: Recomputing the conversation prefix wastes prefill work; retaining every KV cache on GPU limits concurrency.
Method: CachedAttention stores KV state hierarchically, preloads and saves it layer by layer, coordinates fetch/eviction with scheduling, and decouples position encoding from truncation.
Evaluation: workloads=multi-turn LLM conversations; baselines=recompute and GPU-resident KV-cache serving; metrics=TTFT; prefill throughput; end-to-end cost; results=up to 87% lower TTFT; up to 7.8x prefill throughput; up to 70% lower cost
Interpretation: KV persistence is effective when storage movement is overlapped at layer granularity and made scheduler-visible.
Reusable lesson: Persist reusable intermediate state below the accelerator and pipeline its movement with computation.
Applicability: Multi-turn chatbot and agent services with repeated conversation prefixes.
Limits: Benefits depend on prefix reuse, storage bandwidth/capacity, context truncation, and scheduler locality. +
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_cpc_coordinated_page_cache_for_serverless_computing_a8af4792. 확인 범위: official_abstract. 확인한 자료: https://yonsei.elsevierpure.com/en/publications/cpc-coordinated-page-cache-for-serverless-computing/ ; https://doi.org/10.1109/PACT65351.2025.00019 ; https://pact2025.github.io/program. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Bibliographic paper record. +
Verification: official_abstract; confidence=medium.
Canonical title: CrossPrefetch: Accelerating I/O Prefetching for Modern Storage
Question: How can I/O prefetching exploit runtime semantics and remain effective for shared files and remote storage?
Context: OS-only prefetchers see block streams but not application intent; demand I/O can also block useful prefetches.
Method: CrossPrefetch exports runtime state to the OS, separates demand/prefetch paths, tracks fine-grained indices for shared files, and adapts prediction aggressiveness.
Evaluation: workloads=microbenchmarks; macrobenchmarks; real workloads; local and remote storage; baselines=existing OS/runtime prefetchers; metrics=I/O throughput; prefetch accuracy; demand interference; results=1.22–3.7x I/O throughput
Interpretation: Cross-layer semantic hints and isolation between speculative and demand traffic are jointly necessary.
Reusable lesson: Expose application intent to lower layers and prevent speculative work from delaying demand work.
Applicability: Data-intensive runtimes over local or remote file systems.
Limits: Requires OS/runtime integration and sufficiently predictable access patterns; aggressive prefetch can waste bandwidth. +
Verification: official_abstract; confidence=high.
Canonical title: CrossPrefetch: Accelerating I/O Prefetching for Modern Storage
Question: How can I/O prefetching exploit runtime semantics and remain effective for shared files and remote storage?
Context: OS-only prefetchers see block streams but not application intent; demand I/O can also block useful prefetches.
Method: CrossPrefetch exports runtime state to the OS, separates demand/prefetch paths, tracks fine-grained indices for shared files, and adapts prediction aggressiveness.
Evaluation: workloads=microbenchmarks; macrobenchmarks; real workloads; local and remote storage; baselines=existing OS/runtime prefetchers; metrics=I/O throughput; prefetch accuracy; demand interference; results=1.22–3.7x I/O throughput
Interpretation: Cross-layer semantic hints and isolation between speculative and demand traffic are jointly necessary.
Reusable lesson: Expose application intent to lower layers and prevent speculative work from delaying demand work.
Applicability: Data-intensive runtimes over local or remote file systems.
Limits: Requires OS/runtime integration and sufficiently predictable access patterns; aggressive prefetch can waste bandwidth. +
d
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_d2fq_device_direct_fair_queueing_for_nvme_ssds_c60d73e1. 확인 범위: full_text. 확인한 자료: https://www.usenix.org/conference/fast21/presentation/woo ; https://www.usenix.org/system/files/fast21-woo.pdf. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_daac_device_reserved_memory_as_an_eviction_based_file_cache_f0f149f0. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1145/2380403.2380439 ; https://yonsei.elsevierpure.com/en/publications/daac-device-reserved-memory-as-an-eviction-based-file-cache/ ; https://koasas.kaist.ac.kr/handle/10203/175991. 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. +
Verification: arXiv full-text excerpts, DOI metadata, and official acceptance page; confidence=medium.
Canonical title: DeltaZip: Efficient Serving of Multiple Full-Model-Tuned LLMs
Question: How can many fine-tuned LLM variants be served without storing and loading a full weight copy per tenant?
Context: Fine-tuned models share a base lineage but conventional runtimes duplicate weights and serialize costly model swaps.
Method: DeltaZip stores a shared base plus structured-sparse, quantized, optionally lossless deltas and executes custom sparse-batched matrix multiplication in a vLLM-based server.
Evaluation: workloads=fine-tuned Llama-family variants and multi-tenant request traces; baselines=vLLM and conventional model swapping; metrics=delta size, throughput, end-to-end latency, TTFT; results=up to 13x delta compression; 2–12x throughput; 1.6–16x latency/TTFT improvements
Interpretation: Cross-model weight similarity is a serving resource that can be exploited without restricting adaptation to PEFT.
Reusable lesson: Represent related model variants as compact executable deltas, not independent checkpoints.
Applicability: Multi-tenant serving of full- or parameter-efficiently fine-tuned models sharing a base.
Limits: Prompt computation is not accelerated; gains narrow at uniformly high load, and compression needs calibration plus known base lineage. +
Bibliographic paper record. +