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.
s
stem streaming based fpga acceleration for large scale compactions in lsm kv d204166c +
Bibliographic paper record.  +
stem streaming based fpga acceleration for large scale compactions in lsm kv d204166c +
Verification: official_abstract; confidence=high. Question: How can FPGA compaction scale to large LSM runs without materializing oversized intermediate data? Context: Large-scale merge/compaction is CPU- and I/O-intensive, and rigid FPGA pipelines scale poorly across run sizes. Method: STEM uses clean-cut streaming compaction, a multi-unit FPGA pipeline, and dynamic pipeline scheduling. Evaluation: workloads=large-scale LSM compactions; baselines=RocksDB compaction; state-of-the-art FPGA compaction accelerator; metrics=compaction speedup; pipeline utilization; results=27x average and 35x maximum vs RocksDB; 2.09–2.27x vs prior FPGA accelerator Interpretation: Streaming boundaries and elastic pipeline allocation are central to sustaining accelerator utilization. Reusable lesson: Design data-reorganization accelerators around streaming partitions and dynamically scheduled units. Applicability: Large-scale LSM compaction on FPGA-enabled storage servers. Limits: Benefits depend on FPGA resources, host/storage balance, and the evaluated compaction distributions.  +
sti turbocharge nlp inference at the edge via elastic pipelining 99bb4c50 +
Verification: abstract_only; confidence=medium. Canonical title: STI: Turbocharge NLP Inference at the Edge via Elastic Pipelining Question: 메모리가 작은 엣지 장치에서 NLP 모델을 목표 지연 내 실행할 수 있는가? Context: 대형 모델 가중치는 장치 메모리를 넘고 저장소 I/O가 추론을 막는다. Method: 정확도 중요도 기반 샤딩, 탄력적 I/O-연산 파이프라인, 선적재 버퍼를 결합한다. Evaluation: workloads=NLP tasks on two commodity SoCs; CPU and GPU; baselines=prior edge model-loading methods; metrics=memory footprint; latency; accuracy; results=1–2 orders-of-magnitude lower memory with high accuracy. Interpretation: 가중치 중요도와 장치 자원 상태를 함께 이용하면 메모리-지연 절충을 제어할 수 있다. Reusable lesson: 모든 가중치를 동일하게 취급하지 말고 정확도 기여도에 따라 상주시켜라. Applicability: 모바일·엣지 NLP 추론. Limits: 중요도 프로파일과 장치별 파이프라인 튜닝이 필요하다.  +
sti turbocharge nlp inference at the edge via elastic pipelining 99bb4c50 +
Bibliographic paper record.  +
sti turbocharge nlp inference at the edge via elastic pipelining 99bb4c50 +
Verification: abstract_only; confidence=high. Canonical title: STI: Turbocharge NLP Inference at the Edge via Elastic Pipelining Question: 메모리가 작은 엣지 장치에서 NLP 모델을 목표 지연 내 실행할 수 있는가? Context: 대형 모델 가중치는 장치 메모리를 넘고 저장소 I/O가 추론을 막는다. Method: 정확도 중요도 기반 샤딩, 탄력적 I/O-연산 파이프라인, 선적재 버퍼를 결합한다. Evaluation: workloads=NLP tasks on two commodity SoCs; CPU and GPU; baselines=prior edge model-loading methods; metrics=memory footprint; latency; accuracy; results=1–2 orders-of-magnitude lower memory with high accuracy. Interpretation: 가중치 중요도와 장치 자원 상태를 함께 이용하면 메모리-지연 절충을 제어할 수 있다. Reusable lesson: 모든 가중치를 동일하게 취급하지 말고 정확도 기여도에 따라 상주시켜라. Applicability: 모바일·엣지 NLP 추론. Limits: 중요도 프로파일과 장치별 파이프라인 튜닝이 필요하다.  +
swam revisiting swap and oomk for improving application responsiveness on mobile devices 79b3b8a3 +
Verification: full_text; confidence=high. Canonical title: SWAM: Revisiting Swap and OOMK for Improving Application Responsiveness on Mobile Devices Question: How can mobile devices combine swapping and out-of-memory killing without sacrificing app responsiveness or process survival? Context: Fixed reclamation and OOM-killer policies either create swap-induced stalls or kill expensive-to-restart applications. Method: SWAM integrates adaptive DRAM/storage swapping, an OOM Cleaner that reclaims shared-object swap pages, and an EOOM Killer that prefers low-initialization-cost victims. Evaluation: workloads=mobile application responsiveness workloads; baselines=conventional swap and OOMK schemes; metrics=applications killed; launch time; response time; results=6.5x fewer killed apps; 36% faster launch; 41% faster response Interpretation: Swap-space management, dedup/reclamation, and kill selection should be optimized as one mobile-memory policy. Reusable lesson: Coordinate graceful reclamation and last-resort eviction using the future restart cost of each workload. Applicability: Memory-constrained mobile operating systems. Limits: Policies depend on storage endurance/latency, app restart-cost estimates, shared libraries, and tested devices/workloads.  +
swam revisiting swap and oomk for improving application responsiveness on mobile devices 79b3b8a3 +
Verification: full_text; confidence=high. Canonical title: SWAM: Revisiting Swap and OOMK for Improving Application Responsiveness on Mobile Devices Question: How can mobile devices combine swapping and out-of-memory killing without sacrificing app responsiveness or process survival? Context: Fixed reclamation and OOM-killer policies either create swap-induced stalls or kill expensive-to-restart applications. Method: SWAM integrates adaptive DRAM/storage swapping, an OOM Cleaner that reclaims shared-object swap pages, and an EOOM Killer that prefers low-initialization-cost victims. Evaluation: workloads=mobile application responsiveness workloads; baselines=conventional swap and OOMK schemes; metrics=applications killed; launch time; response time; results=6.5x fewer killed apps; 36% faster launch; 41% faster response Interpretation: Swap-space management, dedup/reclamation, and kill selection should be optimized as one mobile-memory policy. Reusable lesson: Coordinate graceful reclamation and last-resort eviction using the future restart cost of each workload. Applicability: Memory-constrained mobile operating systems. Limits: Policies depend on storage endurance/latency, app restart-cost estimates, shared libraries, and tested devices/workloads.  +
swam revisiting swap and oomk for improving application responsiveness on mobile devices 79b3b8a3 +
Bibliographic paper record.  +
symbiosis the art of application and kernel cache cooperation 168e4171 +
Verification: abstract_only; confidence=medium. Canonical title: Symbiosis: The Art of Application and Kernel Cache Cooperation Question: 애플리케이션 캐시와 커널 page cache의 중복·경쟁을 동적으로 조정할 수 있는가? Context: 고정 메모리 분할은 workload 변화와 두 캐시의 중복 때문에 비효율적이다. Method: Symbiosis는 애플리케이션과 커널 캐시 사이의 메모리 비율을 런타임에 조정한다. Evaluation: workloads=LevelDB; WiredTiger; RocksDB; synthetic and real read-heavy workloads; baselines=static cache split; metrics=application performance; results=1.5× average and >5× best-case improvement. Interpretation: 캐시 계층을 독립 튜닝하지 말고 전체 hit·메모리 효율로 공동 제어해야 한다. Reusable lesson: 중복 캐시는 정적 할당보다 피드백 기반 공동 할당이 낫다. Applicability: 자체 캐시를 가진 DB와 OS page cache. Limits: 읽기 중심 평가이며 급격한 변화에서 제어 오버헤드와 안정성이 변수다.  +
symbiosis the art of application and kernel cache cooperation 168e4171 +
Verification: abstract_only; confidence=high. Canonical title: Symbiosis: The Art of Application and Kernel Cache Cooperation Question: 애플리케이션 캐시와 커널 page cache의 중복·경쟁을 동적으로 조정할 수 있는가? Context: 고정 메모리 분할은 workload 변화와 두 캐시의 중복 때문에 비효율적이다. Method: Symbiosis는 애플리케이션과 커널 캐시 사이의 메모리 비율을 런타임에 조정한다. Evaluation: workloads=LevelDB; WiredTiger; RocksDB; synthetic and real read-heavy workloads; baselines=static cache split; metrics=application performance; results=1.5× average and >5× best-case improvement. Interpretation: 캐시 계층을 독립 튜닝하지 말고 전체 hit·메모리 효율로 공동 제어해야 한다. Reusable lesson: 중복 캐시는 정적 할당보다 피드백 기반 공동 할당이 낫다. Applicability: 자체 캐시를 가진 DB와 OS page cache. Limits: 읽기 중심 평가이며 급격한 변화에서 제어 오버헤드와 안정성이 변수다.  +
symbiosis the art of application and kernel cache cooperation 168e4171 +
Bibliographic paper record.  +
t
taming throughput latency tradeoff in llm inference with sarathi serve cf395778 +
Bibliographic paper record.  +
taming throughput latency tradeoff in llm inference with sarathi serve cf395778 +
Verification: full_text; confidence=high. Question: How can LLM serving raise throughput without prefill requests causing decode stalls and tail-latency violations? Context: Prefill and decode have different compute profiles; naive batching creates head-of-line stalls and pipeline bubbles. Method: Sarathi-Serve uses chunked prefills and stall-free scheduling to build uniform hybrid batches and reduce pipeline bubbles. Evaluation: workloads=Mistral-7B on 1xA100; Yi-34B on 2xA100; LLaMA2-70B on 8xA40; Falcon-180B on 8xA100; arXiv summarization trace; baselines=vLLM; metrics=serving capacity; TTFT; TBT; pipeline bubbles; results=2.6x Mistral-7B capacity; 3.7x Yi-34B capacity; 5.6x Falcon-180B capacity Interpretation: Constraining prefill work per iteration makes latency predictable while retaining batch efficiency. Reusable lesson: Split long, bursty stages into schedulable chunks and co-batch them with steady-state work. Applicability: Autoregressive LLM services, especially pipeline-parallel deployments. Limits: Speedups depend on model, GPU topology, arrival trace, chunk size, and chosen TTFT/TBT SLOs.  +
taming throughput latency tradeoff in llm inference with sarathi serve cf395778 +
Verification: full_text; confidence=high. Question: How can LLM serving raise throughput without prefill requests causing decode stalls and tail-latency violations? Context: Prefill and decode have different compute profiles; naive batching creates head-of-line stalls and pipeline bubbles. Method: Sarathi-Serve uses chunked prefills and stall-free scheduling to build uniform hybrid batches and reduce pipeline bubbles. Evaluation: workloads=Mistral-7B on 1xA100; Yi-34B on 2xA100; LLaMA2-70B on 8xA40; Falcon-180B on 8xA100; arXiv summarization trace; baselines=vLLM; metrics=serving capacity; TTFT; TBT; pipeline bubbles; results=2.6x Mistral-7B capacity; 3.7x Yi-34B capacity; 5.6x Falcon-180B capacity Interpretation: Constraining prefill work per iteration makes latency predictable while retaining batch efficiency. Reusable lesson: Split long, bursty stages into schedulable chunks and co-batch them with steady-state work. Applicability: Autoregressive LLM services, especially pipeline-parallel deployments. Limits: Speedups depend on model, GPU topology, arrival trace, chunk size, and chosen TTFT/TBT SLOs.  +
task aware virtual machine scheduling for i o performance f431ccbe +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_task_aware_virtual_machine_scheduling_for_i_o_performance_26a8da47. 확인 범위: full_text. 확인한 자료: https://doi.org/10.1145/1508293.1508308 ; https://yonsei.elsevierpure.com/en/publications/task-aware-virtual-machine-scheduling-for-io-performance/ ; https://www.researchgate.net/publication/221137884_Task-aware_virtual_machine_scheduling_for_IO_performance. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다.  +
task aware virtual machine scheduling for i o performance f431ccbe +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1.  +
technical review 56ee25b0 +
Verification level: official_abstract. Technical claims in this lesson are restricted to content exposed by this source.  +
technical review 5b128265 +
Verification level: official_abstract. Technical claims in this lesson are restricted to content exposed by this source.  +
technical review a case for hardware based demand paging 163a61b2 +
Verification level: official_abstract. Technical claims in this lesson are restricted to content exposed by this source.  +
technical review a fully associative tagless dram cache 8377cc32 +
Verification level: official_abstract. Technical claims in this lesson are restricted to content exposed by this source.  +
원본 주소 "https://s3wiki.yonsei.ac.kr/index.php?title=속성:Evidence_note&oldid=161"