속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
c
Verification: arXiv abstract/full text, DOI metadata, and official SOSP listing; confidence=high.
Canonical title: Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference
Question: How should mobile GPU and NPU execution be combined under a shared memory-bandwidth limit?
Context: Peak accelerator specifications obscure synchronization, operator support, and unified-bandwidth contention in real SoCs.
Method: The study characterizes GPU/NPU behavior and builds HeteroInfer with heterogeneous parallelism and fast synchronization on Snapdragon 8 Gen 3.
Evaluation: workloads=mobile LLM inference on Snapdragon 8 Gen 3; baselines=GPU-only and NPU-only; metrics=inference speed; results=1.34–6.02x speedup
Interpretation: SoC-wide bandwidth and synchronization, not isolated TOPS, determine usable heterogeneous speedup.
Reusable lesson: Profile shared bottlenecks and partition operators across accelerators jointly.
Applicability: On-device LLM inference on unified-memory mobile SoCs.
Limits: Results are tied to one modern Snapdragon platform and its GPU/NPU software stack. +
Bibliographic paper record. +
Verification: arXiv abstract/full text, DOI metadata, and official SOSP listing; confidence=high.
Canonical title: Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference
Question: How should mobile GPU and NPU execution be combined under a shared memory-bandwidth limit?
Context: Peak accelerator specifications obscure synchronization, operator support, and unified-bandwidth contention in real SoCs.
Method: The study characterizes GPU/NPU behavior and builds HeteroInfer with heterogeneous parallelism and fast synchronization on Snapdragon 8 Gen 3.
Evaluation: workloads=mobile LLM inference on Snapdragon 8 Gen 3; baselines=GPU-only and NPU-only; metrics=inference speed; results=1.34–6.02x speedup
Interpretation: SoC-wide bandwidth and synchronization, not isolated TOPS, determine usable heterogeneous speedup.
Reusable lesson: Profile shared bottlenecks and partition operators across accelerators jointly.
Applicability: On-device LLM inference on unified-memory mobile SoCs.
Limits: Results are tied to one modern Snapdragon platform and its GPU/NPU software stack. +
Verification: author-hosted 19-page full paper + official EuroSys 2025 listing; confidence=high.
Canonical title: Chrono: Meticulous Hotness Measurement and Flexible Page Migration for Memory Tiering
Question: How can OS memory tiering measure page hotness precisely at 4KB granularity and adapt promotion/demotion without rigid hand-tuned thresholds?
Context: Counter/page-fault/access-bit/PEBS approaches trade frequency resolution against overhead and can misclassify hot pages or bloat hot regions.
Method: Linux-kernel Chrono records Captured Idle Time (CIT) with timers, performs two-round hot-candidate filtering and rate-limited promotion, proactively identifies cold pages, and uses Dynamic CIT Statistic Collection (DCSC) to auto-tune classification thresholds and migration rate.
Evaluation: Xeon Gold 6348; 64GB DDR4 fast tier + 256GB Optane PM slow tier. Workloads: Pmbench, Graph500, Memcached, Redis, lkp-test. Baselines: Linux NUMA balancing, Auto-Tiering, Multi-Clock, TPP, Memtis. Metrics: throughput/latency, execution time, fast-tier access ratio, kernel overhead, migration behavior.
Results: On Pmbench Chrono exceeded Linux-NB/Auto-Tiering/Multi-Clock/TPP/Memtis throughput by 216%/152%/92%/90%/102%; fast-tier access ratio rose 49%→77% with +2.1 percentage points kernel time vs Linux-NB. Graph500 speedups vs Linux-NB were 2.49×, 2.29×, and 2.05× across 128–256GB working sets. Memcached/Redis also improved overall throughput.
Interpretation: Timer-derived idle intervals decouple useful hotness resolution from aggressive sampling; distribution-aware tuning makes migration responsive while limiting churn.
Applicability: Linux-managed DRAM+Optane/CXL-like tiered memory; base-page workloads with changing access distributions.
Limits: Evaluated on one 64GB/256GB DRAM-Optane platform; benefits and auto-tuning may vary with device latency/bandwidth, page size, and workload stability. Under huge pages Memtis was 1.03× faster in one Graph500 setting. +
Bibliographic paper record. +
Verification: official DOI metadata and EuroSys acceptance listing only; confidence=low.
Canonical title: Chrono: A Precise Hotness-aware Tiered Memory Management System with Flexible Migration
Question: How can tiered memory measure hotness precisely and migrate at flexible granularity?
Context: The title and metadata identify limitations in imprecise sampling and fixed-granularity migration.
Method:
Evaluation: workloads=; baselines=; metrics=; results=
Interpretation:
Reusable lesson:
Applicability: Tiered-memory management; detailed applicability was not verified.
Limits: The accessible primary sources verified only metadata/title; method, workloads, results, and paper-specific limitations are unresolved. +
Verification: author-hosted 19-page full paper + official EuroSys 2025 listing; confidence=high.
Canonical title: Chrono: Meticulous Hotness Measurement and Flexible Page Migration for Memory Tiering.
Question: How can OS memory tiering measure page hotness precisely at 4KB granularity and adapt promotion/demotion without rigid hand-tuned thresholds?
Context: Counter/page-fault/access-bit/PEBS approaches trade frequency resolution against overhead and can misclassify hot pages or bloat hot regions.
Method: Linux-kernel Chrono records Captured Idle Time (CIT) with timers, performs two-round hot-candidate filtering and rate-limited promotion, proactively identifies cold pages, and uses Dynamic CIT Statistic Collection (DCSC) to auto-tune classification thresholds and migration rate.
Evaluation: Xeon Gold 6348; 64GB DDR4 fast tier + 256GB Optane PM slow tier. Workloads: Pmbench, Graph500, Memcached, Redis, lkp-test. Baselines: Linux NUMA balancing, Auto-Tiering, Multi-Clock, TPP, Memtis. Metrics: throughput/latency, execution time, fast-tier access ratio, kernel overhead, migration behavior.
Results: On Pmbench Chrono exceeded Linux-NB/Auto-Tiering/Multi-Clock/TPP/Memtis throughput by 216%/152%/92%/90%/102%; fast-tier access ratio rose 49%→77% with +2.1 percentage points kernel time vs Linux-NB. Graph500 speedups vs Linux-NB were 2.49×, 2.29×, and 2.05× across 128–256GB working sets. Memcached/Redis also improved overall throughput.
Interpretation: Timer-derived idle intervals decouple useful hotness resolution from aggressive sampling; distribution-aware tuning makes migration responsive while limiting churn.
Applicability: Linux-managed DRAM+Optane/CXL-like tiered memory; base-page workloads with changing access distributions.
Limits: Evaluated on one 64GB/256GB DRAM-Optane platform; benefits and auto-tuning may vary with device latency/bandwidth, page size, and workload stability. Under huge pages Memtis was 1.03× faster in one Graph500 setting. +
citadel rethinking memory allocation to safeguard against inter domain rowhammer exploits 8fd39aca +
Verification: official DOI/publisher abstract; confidence=high.
Canonical title: Citadel: Rethinking Memory Allocation to Safeguard Against Inter-Domain Rowhammer Exploits
Question: Can physical memory allocation isolate security domains from cross-domain RowHammer without prohibitive capacity loss?
Context: Guard-row approaches often waste substantial DRAM as the number of isolated domains grows.
Method: Citadel allocates flexible security domains into physically disjoint DRAM regions through a redesigned allocator.
Evaluation: workloads=thousands of memory security domains; baselines=prior guard/isolation allocation; metrics=memory overhead and performance; results=7.2% average overhead, no performance loss, 4–6x less overhead than prior isolation
Interpretation: Row-aware allocator structure can turn physical separation into a scalable isolation primitive.
Reusable lesson: Encode hardware fault geometry directly into allocation domains.
Applicability: Multi-tenant DRAM systems needing inter-domain RowHammer isolation.
Limits: Assumes known DRAM row mapping/geometry and protects inter-domain placement rather than every RowHammer vector. +
citadel rethinking memory allocation to safeguard against inter domain rowhammer exploits 8fd39aca +
Verification: official DOI/publisher abstract; confidence=medium.
Canonical title: Citadel: Rethinking Memory Allocation to Safeguard Against Inter-Domain Rowhammer Exploits
Question: Can physical memory allocation isolate security domains from cross-domain RowHammer without prohibitive capacity loss?
Context: Guard-row approaches often waste substantial DRAM as the number of isolated domains grows.
Method: Citadel allocates flexible security domains into physically disjoint DRAM regions through a redesigned allocator.
Evaluation: workloads=thousands of memory security domains; baselines=prior guard/isolation allocation; metrics=memory overhead and performance; results=7.2% average overhead, no performance loss, 4–6x less overhead than prior isolation
Interpretation: Row-aware allocator structure can turn physical separation into a scalable isolation primitive.
Reusable lesson: Encode hardware fault geometry directly into allocation domains.
Applicability: Multi-tenant DRAM systems needing inter-domain RowHammer isolation.
Limits: Assumes known DRAM row mapping/geometry and protects inter-domain placement rather than every RowHammer vector. +
citadel rethinking memory allocation to safeguard against inter domain rowhammer exploits 8fd39aca +
Bibliographic paper record. +
Bibliographic paper record. +
Verification: abstract_only; confidence=high.
Canonical title: Combining Buffered I/O and Direct I/O in Distributed File Systems
Question: 분산 파일시스템이 요청별로 buffered/direct I/O를 자동 선택할 수 있는가?
Context: 작은 I/O는 버퍼링이 유리하지만 큰·경합 I/O는 direct가 유리해 정적 선택이 어렵다.
Method: AutoIO가 크기·락 경합·메모리 상태를 보고 경로를 동적으로 선택하고 Lustre delayed allocation과 통합한다.
Evaluation: workloads=Lustre client/server workloads; baselines=original Lustre; other distributed file systems; metrics=I/O performance; results=Up to 3× vs Lustre and 13× vs other DFSs.
Interpretation: I/O 경로는 파일 단위가 아니라 순간 요청·자원 상태에 맞춰 선택해야 한다.
Reusable lesson: buffered/direct 선택을 애플리케이션 설정이 아닌 런타임 정책으로 만들라.
Applicability: Lustre·HPC 분산 파일시스템.
Limits: Lustre/HPC 중심이며 임계값과 워크로드 변화에 민감하다. +
Verification: abstract_only; confidence=medium.
Canonical title: Combining Buffered I/O and Direct I/O in Distributed File Systems
Question: 분산 파일시스템이 요청별로 buffered/direct I/O를 자동 선택할 수 있는가?
Context: 작은 I/O는 버퍼링이 유리하지만 큰·경합 I/O는 direct가 유리해 정적 선택이 어렵다.
Method: AutoIO가 크기·락 경합·메모리 상태를 보고 경로를 동적으로 선택하고 Lustre delayed allocation과 통합한다.
Evaluation: workloads=Lustre client/server workloads; baselines=original Lustre; other distributed file systems; metrics=I/O performance; results=Up to 3× vs Lustre and 13× vs other DFSs.
Interpretation: I/O 경로는 파일 단위가 아니라 순간 요청·자원 상태에 맞춰 선택해야 한다.
Reusable lesson: buffered/direct 선택을 애플리케이션 설정이 아닌 런타임 정책으로 만들라.
Applicability: Lustre·HPC 분산 파일시스템.
Limits: Lustre/HPC 중심이며 임계값과 워크로드 변화에 민감하다. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_compressed_memory_swap_for_qos_of_virtualized_embedded_systems_1bbb61cb. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1109/TCE.2012.6311325 ; https://yonsei.elsevierpure.com/en/publications/compressed-memory-swap-for-qos-of-virtualized-embedded-systems/. 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: full_text; confidence=high.
Question: Can LSM-tree compaction policies be expressed as a systematic design space rather than isolated named strategies?
Context: Compaction choices trade write, read, and space amplification, but prior systems mix several design decisions.
Method: The paper decomposes compaction into trigger, data layout, granularity, and data movement, then instantiates and analyzes ten strategies.
Evaluation: workloads=ten instantiated compaction strategies; baselines=representative leveling, tiering, and hybrid policies; metrics=write amplification; write throughput; point lookup; range lookup; space amplification;.; results=12 empirical observations; seven design takeaways; no universal winner
Interpretation: Separating policy primitives reveals which choice causes each tradeoff and enables deliberate hybrid designs.
Reusable lesson: Factor a policy into orthogonal primitives before comparing implementations.
Applicability: LSM key-value stores and compaction-policy design.
Limits: The explored strategies and workloads do not exhaust the design space, and the study does not provide an online adaptive selector. +
Verification: full_text; confidence=high.
Question: Can LSM-tree compaction policies be expressed as a systematic design space rather than isolated named strategies?
Context: Compaction choices trade write, read, and space amplification, but prior systems mix several design decisions.
Method: The paper decomposes compaction into trigger, data layout, granularity, and data movement, then instantiates and analyzes ten strategies.
Evaluation: workloads=ten instantiated compaction strategies; baselines=representative leveling, tiering, and hybrid policies; metrics=write amplification; write throughput; point lookup; range lookup; space amplification;...; results=12 empirical observations; seven design takeaways; no universal winner
Interpretation: Separating policy primitives reveals which choice causes each tradeoff and enables deliberate hybrid designs.
Reusable lesson: Factor a policy into orthogonal primitives before comparing implementations.
Applicability: LSM key-value stores and compaction-policy design.
Limits: The explored strategies and workloads do not exhaust the design space, and the study does not provide an online adaptive selector. +
Bibliographic paper record. +
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_controlling_physical_memory_fragmentation_in_mobile_systems_863971bc. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1145/2754169.2754179 ; https://www.sigplan.org/OpenTOC/ismm15.html ; https://yonsei.elsevierpure.com/en/publications/controlling-physical-memory-fragmentation-in-mobile-systems. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +