속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
a
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 1 of 1. +
Bibliographic paper record. +
Bibliographic paper record. +
Verification: official DOI/publisher abstract; confidence=high.
Canonical title: Asynchrony and GPUs: Bridging this Dichotomy for I/O with AGIO
Question: How can GPU threads issue asynchronous SSD I/O and continue useful work without CPU orchestration?
Context: Synchronous GPU-centric I/O stalls warps; CPU-managed asynchrony sacrifices GPU autonomy and adds coordination.
Method: AGIO decouples initiation/completion, overlaps useful computation, and proactively issues known or data-dependent accesses entirely from the GPU.
Evaluation: workloads=known-pattern and data-dependent GPU I/O workloads; baselines=synchronous GPU-centric I/O; metrics=performance and compute-engine use; results=up to 65% and 32% improvement; <50% compute engines for equal/better performance
Interpretation: A GPU-native completion model turns I/O wait slots into useful compute capacity.
Reusable lesson: Expose nonblocking initiation plus device-local progress and continuation mechanisms.
Applicability: GPU applications with direct SSD access and overlapable computation.
Limits: Requires GPU/SSD integration and enough independent work; proactive access is easier for predictable patterns. +
Verification: official DOI/publisher abstract; confidence=medium.
Canonical title: Asynchrony and GPUs: Bridging this Dichotomy for I/O with AGIO
Question: How can GPU threads issue asynchronous SSD I/O and continue useful work without CPU orchestration?
Context: Synchronous GPU-centric I/O stalls warps; CPU-managed asynchrony sacrifices GPU autonomy and adds coordination.
Method: AGIO decouples initiation/completion, overlaps useful computation, and proactively issues known or data-dependent accesses entirely from the GPU.
Evaluation: workloads=known-pattern and data-dependent GPU I/O workloads; baselines=synchronous GPU-centric I/O; metrics=performance and compute-engine use; results=up to 65% and 32% improvement; <50% compute engines for equal/better performance
Interpretation: A GPU-native completion model turns I/O wait slots into useful compute capacity.
Reusable lesson: Expose nonblocking initiation plus device-local progress and continuation mechanisms.
Applicability: GPU applications with direct SSD access and overlapable computation.
Limits: Requires GPU/SSD integration and enough independent work; proactive access is easier for predictable patterns. +
Bibliographic paper record. +
Verification: abstract_only; confidence=high.
Canonical title: Automatically Reasoning About How Systems Code Uses the CPU Cache
Question: 시스템 바이너리의 CPU cache 사용을 자동으로 추론해 성능·보안 문제를 찾을 수 있는가?
Context: 캐시 footprint·miss·비밀 의존 접근은 수동 측정으로 입력 전반을 설명하기 어렵다.
Method: CFAR가 프로그램 분석과 binary instrumentation으로 memory access를 추출하고 projector로 입력별 cache 동작을 예측한다.
Evaluation: workloads=4 TCP stacks; 7 OpenSSL algorithms; 51 Hyperkernel syscalls; 2 hash tables; baselines=manual cache analysis; metrics=footprint; hits/misses; bug/vulnerability discovery; results=Found performance bugs, vulnerabilities, and third-party cache impacts; no aggregate numeric speedup in abstract.
Interpretation: 저수준 접근을 입력 공간의 함수로 요약하면 cache 문제를 자동 탐색할 수 있다.
Reusable lesson: cache 분석은 단일 trace가 아니라 입력→footprint/miss의 모델로 표현하라.
Applicability: OS·network stack·crypto 코드 분석.
Limits: 분석 가능한 코드·입력 모델·instrumentation coverage와 projector 정확도에 의존한다. +
Verification: abstract_only; confidence=medium.
Canonical title: Automatically Reasoning About How Systems Code Uses the CPU Cache
Question: 시스템 바이너리의 CPU cache 사용을 자동으로 추론해 성능·보안 문제를 찾을 수 있는가?
Context: 캐시 footprint·miss·비밀 의존 접근은 수동 측정으로 입력 전반을 설명하기 어렵다.
Method: CFAR가 프로그램 분석과 binary instrumentation으로 memory access를 추출하고 projector로 입력별 cache 동작을 예측한다.
Evaluation: workloads=4 TCP stacks; 7 OpenSSL algorithms; 51 Hyperkernel syscalls; 2 hash tables; baselines=manual cache analysis; metrics=footprint; hits/misses; bug/vulnerability discovery; results=Found performance bugs, vulnerabilities, and third-party cache impacts; no aggregate numeric speedup in abstract.
Interpretation: 저수준 접근을 입력 공간의 함수로 요약하면 cache 문제를 자동 탐색할 수 있다.
Reusable lesson: cache 분석은 단일 trace가 아니라 입력→footprint/miss의 모델로 표현하라.
Applicability: OS·network stack·crypto 코드 분석.
Limits: 분석 가능한 코드·입력 모델·instrumentation coverage와 projector 정확도에 의존한다. +
Bibliographic paper record. +
Verification: official_abstract; confidence=medium.
Question: Can a buffer manager serve cache-miss reads immediately when every frame is dirty?
Context: Read-after-writeback (RAW) stalls the foreground read until a dirty victim page is flushed to flash.
Method: WAR copies the dirty LRU-tail page to a DRAM staging area, immediately reuses the frame for the read, and flushes the copy asynchronously.
Evaluation: workloads=MySQL/InnoDB; Zero; baselines=RAW read-after-writeback; metrics=transaction throughput; read latency; stall time; results=up to 2.9x transaction throughput
Interpretation: A cheap memory copy can remove a long storage dependency from the foreground path.
Reusable lesson: Use bounded staging to decouple latency-critical allocation from slow durable writeback.
Applicability: Flash-backed databases and buffer caches under dirty-page pressure.
Limits: Requires extra DRAM and careful ordering/recovery; value falls when dirty-victim stalls are rare. +
Verification: official_abstract; confidence=high.
Question: Can a buffer manager serve cache-miss reads immediately when every frame is dirty?
Context: Read-after-writeback (RAW) stalls the foreground read until a dirty victim page is flushed to flash.
Method: WAR copies the dirty LRU-tail page to a DRAM staging area, immediately reuses the frame for the read, and flushes the copy asynchronously.
Evaluation: workloads=MySQL/InnoDB; Zero; baselines=RAW read-after-writeback; metrics=transaction throughput; read latency; stall time; results=up to 2.9x transaction throughput
Interpretation: A cheap memory copy can remove a long storage dependency from the foreground path.
Reusable lesson: Use bounded staging to decouple latency-critical allocation from slow durable writeback.
Applicability: Flash-backed databases and buffer caches under dirty-page pressure.
Limits: Requires extra DRAM and careful ordering/recovery; value falls when dirty-victim stalls are rare. +
aware automate workload autoscaling with reinforcement learning in production cloud systems cb18da5e +
Verification: abstract_only; confidence=medium.
Canonical title: AWARE: Automate Workload Autoscaling with Reinforcement Learning in Production Cloud Systems
Question: RL 오토스케일러를 새 워크로드에 빠르고 안전하게 적용할 수 있는가?
Context: RL은 학습 비용·초기 안전성·워크로드 변화 때문에 운영 도입이 어렵다.
Method: 메타러닝과 안전한 부트스트래핑으로 에이전트를 자동 생성·적응시킨다.
Evaluation: workloads=production cloud serving and training workloads; baselines=transfer learning; metrics=adaptation speed; reward; utilization; SLO violations; results=5.5× faster adaptation; +47.5% CPU/+39.2% memory utilization; 16.9× fewer SLO violations in training.
Interpretation: 운영 RL에는 정책 자체보다 안전한 초기화와 빠른 전이가 중요하다.
Reusable lesson: 온라인 제어 RL은 메타학습과 안전 제약을 함께 설계하라.
Applicability: 클라우드 서비스 오토스케일링.
Limits: 보상 설계·시뮬레이터 전이·운영 안전성 검증에 의존한다. +
aware automate workload autoscaling with reinforcement learning in production cloud systems cb18da5e +
Verification: abstract_only; confidence=high.
Canonical title: AWARE: Automate Workload Autoscaling with Reinforcement Learning in Production Cloud Systems
Question: RL 오토스케일러를 새 워크로드에 빠르고 안전하게 적용할 수 있는가?
Context: RL은 학습 비용·초기 안전성·워크로드 변화 때문에 운영 도입이 어렵다.
Method: 메타러닝과 안전한 부트스트래핑으로 에이전트를 자동 생성·적응시킨다.
Evaluation: workloads=production cloud serving and training workloads; baselines=transfer learning; metrics=adaptation speed; reward; utilization; SLO violations; results=5.5× faster adaptation; +47.5% CPU/+39.2% memory utilization; 16.9× fewer SLO violations in training.
Interpretation: 운영 RL에는 정책 자체보다 안전한 초기화와 빠른 전이가 중요하다.
Reusable lesson: 온라인 제어 RL은 메타학습과 안전 제약을 함께 설계하라.
Applicability: 클라우드 서비스 오토스케일링.
Limits: 보상 설계·시뮬레이터 전이·운영 안전성 검증에 의존한다. +
aware automate workload autoscaling with reinforcement learning in production cloud systems cb18da5e +
Bibliographic paper record. +
b
Yonsei University Computer Systems Laboratory publication list; authors: Jeonghyun Lee, Minwoo Ahn, Jisung Park, Jinkyu Jeong. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_beyond_static_policies_dynamic_prac_for_balanced_and_efficient_rowhammer_mitiga_185f5688. 확인 범위: metadata_only. 확인한 자료: https://www.islped.org/2026/final-program/ ; https://jisung-park.github.io/publications ; https://cslab.yonsei.ac.kr/publications. 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. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_bperf_bcoz_on_cpu_off_cpu_7ecf409a. 확인 범위: partial_source. 확인한 자료: https://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE12102543 ; https://cslab.yonsei.ac.kr/publications. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Verification: full_text; confidence=high.
Canonical title: BreakHammer: Enhancing RowHammer Mitigations by Carefully Throttling Suspect Threads
Question: Can RowHammer defenses reduce their own performance/energy cost by slowing the threads that trigger preventive actions?
Context: Existing mitigations protect rows globally, so one aggressive or malicious thread can impose refresh/repair overhead on all workloads.
Method: BreakHammer monitors preventive-action attribution, identifies suspect threads, and throttles their memory bandwidth.
Evaluation: workloads=eight RowHammer mitigation mechanisms; baselines=each mitigation without BreakHammer; metrics=performance; DRAM energy; fairness; area; results=improves all evaluated mitigations; near-zero area overhead
Interpretation: Attributing defense cost to its trigger lets the system contain both attacks and collateral overhead.
Reusable lesson: Add per-principal attribution and feedback throttling around expensive shared defenses.
Applicability: DRAM systems deploying preventive RowHammer mitigations.
Limits: Relies on correct culprit identification and simulated mitigation models; benign high-activity threads can be throttled. +
Verification: full_text; confidence=high.
Canonical title: BreakHammer: Enabling Scalable and Low Overhead RowHammer Mitigations via Throttling Preventive Action Triggering Threads
Question: Can RowHammer defenses reduce their own performance/energy cost by slowing the threads that trigger preventive actions?
Context: Existing mitigations protect rows globally, so one aggressive or malicious thread can impose refresh/repair overhead on all workloads.
Method: BreakHammer monitors preventive-action attribution, identifies suspect threads, and throttles their memory bandwidth.
Evaluation: workloads=eight RowHammer mitigation mechanisms; baselines=each mitigation without BreakHammer; metrics=performance; DRAM energy; fairness; area; results=improves all evaluated mitigations; near-zero area overhead
Interpretation: Attributing defense cost to its trigger lets the system contain both attacks and collateral overhead.
Reusable lesson: Add per-principal attribution and feedback throttling around expensive shared defenses.
Applicability: DRAM systems deploying preventive RowHammer mitigations.
Limits: Relies on correct culprit identification and simulated mitigation models; benign high-activity threads can be throttled. +