본문으로 이동

속성:Evidence note

S3 연구 메모리

Text

자료에서 확인할 수 있는 점이나 한계를 적습니다.

( | ) (20 | 50 | 100 | 250 | 500) 보기
이 속성을 사용하는 문서 20개를 보여줍니다.
f
following the data not the function rethinking function orchestration in serverless computing 5edeca75 +
Bibliographic paper record.  +
fully harnessing the performance potential of dram less mobile flash storage ff995292 +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1.  +
fully harnessing the performance potential of dram less mobile flash storage ff995292 +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_fully_harnessing_the_performance_potential_of_dram_less_mobile_flash_storage_d4449ca9. 확인 범위: full_text. 확인한 자료: https://www.msstconference.org/2024/ ; https://www.msstconference.org/MSST-history/2024/Papers/msst24-7.3.pdf. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다.  +
g
gpu c88d9a4d +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_gpu_e130421c. 확인 범위: metadata_only. 확인한 자료: https://www.riss.kr/search/Search.do?colName=re_a_kor&isDetailSearch=Y&queryText=znCreator%2C%EC%95%88%EB%AF%BC%EC%9A%B0%28Minwoo+Ahn%29&searchGubun=true ; https://cslab.yonsei.ac.kr/publications. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다.  +
gpu c88d9a4d +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility.  +
gqa training generalized multi query transformer models from multi head checkpoints a6839ee2 +
Bibliographic paper record.  +
gqa training generalized multi query transformer models from multi head checkpoints a6839ee2 +
Verification: abstract_only; confidence=medium. Canonical title: GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints Question: MHA 품질과 MQA 디코딩 속도 사이의 절충을 기존 체크포인트에서 얻을 수 있는가? Context: MQA는 KV 캐시를 줄이지만 품질 저하가 생길 수 있고 MHA는 추론 메모리가 크다. Method: KV head를 중간 개수의 그룹으로 묶는 GQA와 MHA 체크포인트 업트레이닝을 제안한다. Evaluation: workloads=T5-family language tasks; baselines=MHA; MQA; metrics=quality; inference speed; uptraining compute; results=Near-MHA quality, near-MQA speed using 5% pretraining compute. Interpretation: KV head 공유 정도는 이진 선택이 아니라 연속적 시스템-품질 설계 변수다. Reusable lesson: 기존 체크포인트를 저비용 업트레이닝해 추론 구조를 바꿀 수 있다. Applicability: Transformer 생성 추론과 KV 캐시 절감. Limits: 그룹 수 선택과 5% 추가 학습비가 필요하며 모델군별 일반화가 제한될 수 있다.  +
gqa training generalized multi query transformer models from multi head checkpoints a6839ee2 +
Verification: abstract_only; confidence=high. Canonical title: GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints Question: MHA 품질과 MQA 디코딩 속도 사이의 절충을 기존 체크포인트에서 얻을 수 있는가? Context: MQA는 KV 캐시를 줄이지만 품질 저하가 생길 수 있고 MHA는 추론 메모리가 크다. Method: KV head를 중간 개수의 그룹으로 묶는 GQA와 MHA 체크포인트 업트레이닝을 제안한다. Evaluation: workloads=T5-family language tasks; baselines=MHA; MQA; metrics=quality; inference speed; uptraining compute; results=Near-MHA quality, near-MQA speed using 5% pretraining compute. Interpretation: KV head 공유 정도는 이진 선택이 아니라 연속적 시스템-품질 설계 변수다. Reusable lesson: 기존 체크포인트를 저비용 업트레이닝해 추론 구조를 바꿀 수 있다. Applicability: Transformer 생성 추론과 KV 캐시 절감. Limits: 그룹 수 선택과 5% 추가 학습비가 필요하며 모델군별 일반화가 제한될 수 있다.  +
group based memory oversubscription for virtualized clouds a7129dfa +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1.  +
group based memory oversubscription for virtualized clouds a7129dfa +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_group_based_memory_oversubscription_for_virtualized_clouds_c7cc40cd. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1016/j.jpdc.2014.01.001 ; https://www.sciencedirect.com/science/article/pii/S0743731514000033 ; https://yonsei.elsevierpure.com/en/publications/group-based-memory-oversubscription-for-virtualized-clouds/. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다.  +
h
halfmoon log optimal fault tolerant stateful serverless computing e44c470a +
Bibliographic paper record.  +
halfmoon log optimal fault tolerant stateful serverless computing e44c470a +
Verification: abstract_only; confidence=medium. Canonical title: Halfmoon: Log-Optimal Fault-Tolerant Stateful Serverless Computing Question: stateful serverless의 장애 허용 로그 비용을 작업별 최솟값으로 줄일 수 있는가? Context: 모든 상태 읽기·쓰기를 동일하게 기록하면 불필요한 로그와 지연이 발생한다. Method: 읽기 또는 쓰기만 기록하는 두 비대칭 log-optimal 프로토콜과 선택 기준, 무중단 전환을 제공한다. Evaluation: workloads=stateful serverless workloads; baselines=Boki; metrics=latency; logging overhead; results=20–40% lower latency; 1.5–4× lower logging overhead. Interpretation: 함수의 외부 상태 접근 패턴에 맞춰 복구 로그 방향을 선택해야 한다. Reusable lesson: 복구 프로토콜은 read/write 대칭 가정을 버리고 작업 유형별로 특화하라. Applicability: fault-tolerant stateful FaaS. Limits: 외부 상태 의미론과 workload 분류, 프로토콜 전환 정확성에 의존한다.  +
halfmoon log optimal fault tolerant stateful serverless computing e44c470a +
Verification: abstract_only; confidence=high. Canonical title: Halfmoon: Log-Optimal Fault-Tolerant Stateful Serverless Computing Question: stateful serverless의 장애 허용 로그 비용을 작업별 최솟값으로 줄일 수 있는가? Context: 모든 상태 읽기·쓰기를 동일하게 기록하면 불필요한 로그와 지연이 발생한다. Method: 읽기 또는 쓰기만 기록하는 두 비대칭 log-optimal 프로토콜과 선택 기준, 무중단 전환을 제공한다. Evaluation: workloads=stateful serverless workloads; baselines=Boki; metrics=latency; logging overhead; results=20–40% lower latency; 1.5–4× lower logging overhead. Interpretation: 함수의 외부 상태 접근 패턴에 맞춰 복구 로그 방향을 선택해야 한다. Reusable lesson: 복구 프로토콜은 read/write 대칭 가정을 버리고 작업 유형별로 특화하라. Applicability: fault-tolerant stateful FaaS. Limits: 외부 상태 의미론과 workload 분류, 프로토콜 전환 정확성에 의존한다.  +
harvesting memory bound cpu stall cycles in software with msh 5866da08 +
Bibliographic paper record.  +
harvesting memory bound cpu stall cycles in software with msh 5866da08 +
Verification: official_abstract; confidence=high. Question: Can software safely run useful work during memory-bound CPU stalls when SMT violates latency SLOs? Context: SMT harvests stalls but offers coarse concurrency control and can interfere with latency-critical work. Method: MSH combines profiling, program analysis, binary instrumentation, and a runtime scheduler with configurable harvesting concurrency. Evaluation: workloads=memory-bound applications; names not enumerated in abstract; baselines=SMT; no harvesting; metrics=harvesting throughput; latency SLO compliance; results=up to 72% of SMT harvesting throughput under SLOs where SMT is unusable Interpretation: Explicit software yield points trade some peak harvest for much tighter interference control. Reusable lesson: Expose resource-stall opportunities to a policy-controlled runtime when hardware sharing is too coarse. Applicability: Memory-bound server workloads with latency SLOs and spare CPU cycles. Limits: Requires analyzable/instrumentable binaries and enough predictable memory stalls; abstract does not enumerate all workloads.  +
harvesting memory bound cpu stall cycles in software with msh 5866da08 +
Verification: official_abstract; confidence=medium. Question: Can software safely run useful work during memory-bound CPU stalls when SMT violates latency SLOs? Context: SMT harvests stalls but offers coarse concurrency control and can interfere with latency-critical work. Method: MSH combines profiling, program analysis, binary instrumentation, and a runtime scheduler with configurable harvesting concurrency. Evaluation: workloads=memory-bound applications; names not enumerated in abstract; baselines=SMT; no harvesting; metrics=harvesting throughput; latency SLO compliance; results=up to 72% of SMT harvesting throughput under SLOs where SMT is unusable Interpretation: Explicit software yield points trade some peak harvest for much tighter interference control. Reusable lesson: Expose resource-stall opportunities to a policy-controlled runtime when hardware sharing is too coarse. Applicability: Memory-bound server workloads with latency SLOs and spare CPU cycles. Limits: Requires analyzable/instrumentable binaries and enough predictable memory stalls; abstract does not enumerate all workloads.  +
hotrap hot record retention and promotion for lsm trees with tiered storage 849aa350 +
Verification: official USENIX page and abstract; confidence=medium. Canonical title: HotRAP: Hot Record Retention and Promotion for LSM-trees with Tiered Storage Question: How can an LSM store promote hot records without scanning or rewriting large sorted runs? Context: Coarse promotion wastes I/O and space when only a few keys in an on-disk structure are hot. Method: HotRAP adds a fine-grained on-disk hot-key tracker and two pathways for promoting selected records into RocksDB's faster levels. Evaluation: workloads=YCSB and Twitter traces; baselines=second-best evaluated LSM optimization; metrics=throughput; results=up to 1.6x on balanced YCSB and 1.5x on Twitter Interpretation: Record-level popularity can guide selective LSM reshaping more efficiently than run-level movement. Reusable lesson: Track and move the smallest independently valuable hot unit. Applicability: LSM-tree stores with skewed key popularity. Limits: Benefits depend on hot-key skew and tracker/promotion overhead.  +
hotrap hot record retention and promotion for lsm trees with tiered storage 849aa350 +
Bibliographic paper record.  +
hotrap hot record retention and promotion for lsm trees with tiered storage 849aa350 +
Verification: official USENIX page and abstract; confidence=high. Canonical title: HotRAP: Hot-Record-Aware Promotion for LSM-Tree Key-Value Stores Question: How can an LSM store promote hot records without scanning or rewriting large sorted runs? Context: Coarse promotion wastes I/O and space when only a few keys in an on-disk structure are hot. Method: HotRAP adds a fine-grained on-disk hot-key tracker and two pathways for promoting selected records into RocksDB's faster levels. Evaluation: workloads=YCSB and Twitter traces; baselines=second-best evaluated LSM optimization; metrics=throughput; results=up to 1.6x on balanced YCSB and 1.5x on Twitter Interpretation: Record-level popularity can guide selective LSM reshaping more efficiently than run-level movement. Reusable lesson: Track and move the smallest independently valuable hot unit. Applicability: LSM-tree stores with skewed key popularity. Limits: Benefits depend on hot-key skew and tracker/promotion overhead.  +
how storage i o affects user perceived latency in mobile apps b59ccb75 +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1.  +
원본 주소 "https://s3wiki.yonsei.ac.kr/index.php?title=속성:Evidence_note&oldid=161"