본문으로 이동

속성:Context

S3 연구 메모리

Text

하드웨어, 부하, 버전, 규모처럼 결과에 영향을 줄 수 있는 조건을 적습니다.

( | ) (20 | 50 | 100 | 250 | 500) 보기
이 속성을 사용하는 문서 20개를 보여줍니다.
d
dlora dynamically orchestrating requests and adapters for lora llm serving 3827bfa3 +
Venue: OSDI. Year: 2024. Always merging adapters is slow to switch; never merging them loses kernel efficiency and can create imbalance. Verification: full_text; confidence=high.  +
dram 8d56adf1 +
Publication scope: domestic. Lab publication metadata: 1. 노재선, 정진규, "DRAM 없는 모바일 플래시 스토리지에서의 유효 페이지 추적 기법 연구", 정보과학회논문지, vol.52, no.5, pp. 357-362, 2025 2. 노재선, 정진규, "DRAM 없는 모바일 플래시 스토리지에서의 유효 페이지 추적 기법 연구", 2023 한국소프트웨어종합학술대회 (KSC), 2023 (최우수논문상) Verification level: official_abstract. Source detail: KCI 공식 논문 페이지와 연구실 출판 목록. 원문 전체가 아닌 공식 초록에서 검증했으므로 전체 workload 구성과 부가 비용은 미검증이다.  +
dram architecture for efficient data lifetime management f27c8820 +
Publication scope: international. Lab publication metadata: 1. Yongjun Lee, Yunkeuk Kim, Jinkyu Jeong, and Jae W. Lee, "DRAM Architecture for Efficient Data Lifetime Management," IEICE Electronics Express, vol. 14, no. 10, Article 20170309, 2017 Verification level: metadata_only. Sources checked: https://www.jstage.jst.go.jp/article/elex/advpub/0/advpub_14.20170309/_article/-char/en https://www.jstage.jst.go.jp/article/elex/14/10/14_14.20170309/_pdf  +
dwkv collinear slo confound +
초기 workload에는 premium/interactive/batch마다 하나의 slo_budget만 존재했다. tier value, deadline budget, max_tokens가 사실상 동일한 세 triple로 묶였다.  +
dwkv doomed chase domino +
실행 코드와 simulator 모두 raw laxity가 0 이하일 때 기본 `chase` branch가 urgency floor를 사용했다. 그러면 doomed request가 healthy request보다 최대 약 60배 높은 urgency를 받는다.  +
dwkv failed run fail closed +
archive에는 incomplete sweep, HTTP 400 request failure, seed 일부만 존재하는 control, server startup-timeout repair cell이 포함되어 있다.  +
dwkv kv saturation unproven +
GPU harness는 `max_num_seqs=16`에 도달했지만 retained B200 telemetry의 observed KV use는 0.4–1.9%였고, simulator는 full potential footprint를 예약하며 continuous-batch latency coupling, KV growth, token-budget contention, recomputation을 모델링하지 않았다.  +
dwkv legacy patch not repro +
patch는 early vLLM 0.21.0 prototype의 4개 source file 변경을 담지만, 이후 doomed-request handling, marginal-risk queue, experiment schema, correctness fixes, current policy set이 반영되지 않았다.  +
dwkv metric mismatch +
DW-KV score는 `tier_value × urgency / kv_footprint`로 value per KV byte를 최적화하도록 설계되었지만, 기존 SPEC의 primary target은 premium SLO miss 0%였다.  +
dwkv prompt seed bug +
Qwen2.5-72B-AWQ, 2×B200 online Poisson-arrival 실험. 수정 전 harness는 prompt pool을 global unseeded random으로 섞고 `prompt_pool[i]`를 사용했다.  +
dwkv score code contract +
simulator의 `score()`와 `urgency()` 및 `vllm/v1/core/sched/request_queue.py`의 `_score_debug()`가 정책 차이를 계산한다. queue는 time-dependent urgency 때문에 access마다 re-score한다.  +
e
efficient footprint caching for tagless dram caches 9317d60c +
Publication scope: international. Lab publication metadata: 1. Hakbeom Jang, Yongjun Lee, Jongwon Kim, Youngsok Kim, Jangwoo Kim, Jinkyu Jeong, and Jae W. Lee, "Efficient Footprint Caching for Tagless DRAM Caches," in Proceedings of The 22nd IEEE International Symposium on High-Performance Computer Architecture (HPCA-22), Barcelona, Spain, March 2016. Verification level: official_abstract. Sources checked: https://doi.org/10.1109/HPCA.2016.7446068 https://yonsei.elsevierpure.com/en/publications/efficient-footprint-caching-for-tagless-dram-caches/  +
efficient function call tracing with link time binary rewriting for ce devices 0fd13b41 +
Publication scope: international. Lab publication metadata: 1. Bon-Keun Seo, Jinkyu Jeong, Joonwon Lee, and Euiseong Seo, "Efficient Function Call Tracing with Link-Time Binary Rewriting for CE Devices," IEEE Transactions on Consumer Electronics, vol. 59, no. 4, pp. 892-900, Nov. 2013 Verification level: official_abstract. Sources checked: https://doi.org/10.1109/TCE.2013.6689704 https://yonsei.elsevierpure.com/en/publications/efficient-function-call-tracing-with-link-time-binary-rewriting-f/  +
efficient hybrid polling for ultra low latency storage devices e753019a +
Publication scope: international. Lab publication metadata: 1. Gyusun Lee, Seokha Shin, Jinkyu Jeong, "Efficient Hybrid Polling for Ultra-Low Latency Storage Devices," Journal of Systems Architecture, Volume 122, January 2022, 102338 Verification level: full_text. Sources checked: https://www.sciencedirect.com/science/article/pii/S1383762121002319 https://yonsei.elsevierpure.com/en/publications/efficient-hybrid-polling-for-ultra-low-latency-storage-devices/ https://doi.org/10.1016/j.sysarc.2021.102338  +
efficient memory deduplication for mobile smart devices fa9b72e9 +
Publication scope: international. Lab publication metadata: 1. Sung-Hun Kim, Jinkyu Jeong, and Joonwon Lee, "Efficient Memory Deduplication for Mobile Smart Devices," in Proceedings of IEEE International Conference on Consumer Electronics, Las Vegas, NV, USA, January, 2014 Verification level: official_abstract. Sources checked: https://doi.org/10.1109/ICCE.2014.6775893 https://yonsei.elsevierpure.com/en/publications/efficient-memory-deduplicati-on-for-mobile-smart-devices/  +
efficient memory overcommitment for i o passthrough enabled vms via fine grained page meta data d32f4845 +
Venue: USENIX ATC. Year: 2023. ballooning은 느리고 passthrough 장치의 DMA 때문에 페이지 회수가 위험하다. Verification: abstract_only; confidence=high.  +
emt an os framework for new memory translation architectures 6d3356f6 +
Venue: OSDI. Year: 2025. New radix/hash and hardware-assisted translation designs conflict with assumptions embedded throughout modern VM code. Verification: official USENIX page and abstract; confidence=high.  +
enabling high performance and secure userspace nvm file systems with the trio architecture d59d00b6 +
Venue: SOSP. Year: 2023. 직접 NVM 접근은 빠르지만 메타데이터 위조와 커널 검증 비용 문제가 있다. Verification: abstract_only; confidence=high.  +
energy efficient scheduling of real time tasks on multicore processors a70b387a +
Publication scope: international. Lab publication metadata: 1. Euiseong Seo, Jinkyu Jeong, Seonyeong Park, and Joonwon Lee, "Energy Efficient Scheduling of Real-Time Tasks on Multicore Processors," IEEE Transactions on Parallel and Distributed Systems, Vol. 19, No. 11, pp. 1540-1552, November 2008 Verification level: official_abstract. Sources checked: https://doi.org/10.1109/TPDS.2008.104 https://yonsei.elsevierpure.com/en/publications/energy-efficient-scheduling-of-real-time-tasks-on-multicore-proce/  +
enhancing network i o performance for a virtualized hadoop cluster d869aae7 +
Publication scope: international. Lab publication metadata: 1. Jinkyu Jeong, Dong Hoon Choi, and Heeseung Jo, "Enhancing Network I/O Performance for a Virtualized Hadoop Cluster," Concurrency and Computation Practice and Experience, vol. 29. no. 8, e3974, April 2017 Verification level: official_abstract. Sources checked: https://doi.org/10.1002/cpe.3974 https://yonsei.elsevierpure.com/en/publications/enhancing-network-io-performance-for-a-virtualized-hadoop-cluster/  +
원본 주소 "https://s3wiki.yonsei.ac.kr/index.php?title=속성:Context&oldid=146"