속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
t
Verification: abstract_only; confidence=high.
Canonical title: Tectonic-Shift: A Composite Storage Fabric for Large-Scale ML Training
Question: 대규모 ML 학습 저장소의 전력과 I/O 효율을 함께 높일 수 있는가?
Context: HDD 기반 Tectonic은 학습 읽기 부하를 위해 과도한 용량·전력을 요구한다.
Method: Shift 플래시 계층과 데이터셋 명세에서 미래 접근을 추론하는 애플리케이션 인지 캐시를 결합한다.
Evaluation: workloads=petabyte-scale production ML cluster; baselines=LRU flash cache; metrics=I/O absorption; power; results=1.51–3.28× I/O absorption; 29% lower power.
Interpretation: 학습 데이터 접근의 예측 가능성을 저장 계층 정책에 노출하면 플래시 효율이 커진다.
Reusable lesson: 일반 LRU 대신 애플리케이션 의미를 캐시 정책에 활용하라.
Applicability: 대규모 반복형 ML 데이터 로딩.
Limits: Meta 학습 트레이스와 예측 가능한 데이터셋 순서를 전제한다. +
Verification: official_abstract; confidence=high.
Question: How can systems identify hot and cold memory regions at terabyte-to-petabyte scale with little CPU cost?
Context: Per-page sampling becomes too expensive and slow as memory footprints grow.
Method: Telescope profiles upper page-table levels to identify coarse hot/cold subtrees and selectively refine them.
Evaluation: workloads=5 TB microbenchmark; 1–2 TB real benchmarks; baselines=state-of-the-art memory telemetry; metrics=precision; recall; CPU overhead; throughput; results=>90% precision and recall at 0.9% of one CPU; 5.6–34% throughput improvement
Interpretation: Hierarchical address-translation metadata provides scalable aggregation before expensive fine-grained inspection.
Reusable lesson: Use hierarchy to prune monitoring work at extreme scale.
Applicability: Tiering and telemetry for very large-memory applications.
Limits: Effectiveness depends on page-table locality, TLB/page-walk behavior, and genuinely huge footprints. +
Verification: official_abstract; confidence=medium.
Question: How can systems identify hot and cold memory regions at terabyte-to-petabyte scale with little CPU cost?
Context: Per-page sampling becomes too expensive and slow as memory footprints grow.
Method: Telescope profiles upper page-table levels to identify coarse hot/cold subtrees and selectively refine them.
Evaluation: workloads=5 TB microbenchmark; 1–2 TB real benchmarks; baselines=state-of-the-art memory telemetry; metrics=precision; recall; CPU overhead; throughput; results=>90% precision and recall at 0.9% of one CPU; 5.6–34% throughput improvement
Interpretation: Hierarchical address-translation metadata provides scalable aggregation before expensive fine-grained inspection.
Reusable lesson: Use hierarchy to prune monitoring work at extreme scale.
Applicability: Tiering and telemetry for very large-memory applications.
Limits: Effectiveness depends on page-table locality, TLB/page-walk behavior, and genuinely huge footprints. +
Bibliographic paper record. +
the cost of dynamic reasoning demystifying ai agents and test time scaling from an ai infrastruc 6d6dff58 +
Bibliographic paper record. +
the cost of dynamic reasoning demystifying ai agents and test time scaling from an ai infrastruc 6d6dff58 +
Verification: arXiv metadata and abstract; confidence=medium.
Canonical title: The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective
Question: What system cost is introduced when LLM inference dynamically expands reasoning or agentic work?
Context: Accuracy-oriented techniques such as few-shot prompting, reflection, and parallel reasoning change work per request and its variance.
Method: The study profiles multiple agent and test-time-scaling designs across resource use, latency, energy, and datacenter power.
Evaluation: workloads=agent designs and test-time-scaling configurations; baselines=few-shot, varying reflection depth, and parallel reasoning configurations; metrics=latency, latency variance, resource use, energy, and datacenter power; results=qualitative diminishing returns and rising systems cost
Interpretation: Reasoning quality must be optimized jointly with tail latency, energy, and capacity rather than token count alone.
Reusable lesson: Model dynamic reasoning as a variable-cost systems workload with explicit stopping and resource budgets.
Applicability: Agentic and test-time-compute LLM services.
Limits: Only the arXiv abstract was verified; model suite, hardware, and exact numerical findings were not extracted. +
the cost of dynamic reasoning demystifying ai agents and test time scaling from an ai infrastruc 6d6dff58 +
Verification: arXiv metadata and abstract; confidence=medium.
Canonical title: The Cost of Dynamic Reasoning: A Systematic Characterization of Large Language Model Inference
Question: What system cost is introduced when LLM inference dynamically expands reasoning or agentic work?
Context: Accuracy-oriented techniques such as few-shot prompting, reflection, and parallel reasoning change work per request and its variance.
Method: The study profiles multiple agent and test-time-scaling designs across resource use, latency, energy, and datacenter power.
Evaluation: workloads=agent designs and test-time-scaling configurations; baselines=few-shot, varying reflection depth, and parallel reasoning configurations; metrics=latency, latency variance, resource use, energy, and datacenter power; results=qualitative diminishing returns and rising systems cost
Interpretation: Reasoning quality must be optimized jointly with tail latency, energy, and capacity rather than token count alone.
Reusable lesson: Model dynamic reasoning as a variable-cost systems workload with explicit stopping and resource budgets.
Applicability: Agentic and test-time-compute LLM services.
Limits: Only the arXiv abstract was verified; model suite, hardware, and exact numerical findings were not extracted. +
Verification: abstract_only; confidence=medium.
Canonical title: The Design and Implementation of a Capacity-Variant Storage System
Question: SSD가 수명·성능을 위해 실행 중 용량을 탄력적으로 줄일 수 있는가?
Context: 고정 over-provisioning은 workload 변화에 따라 용량·GC·수명 절충을 최적화하지 못한다.
Method: CV-SSD, 탄력 로그 구조 CV-FS, CV-manager가 용량을 점진적으로 조정한다.
Evaluation: workloads=synthetic and real workloads; baselines=fixed-capacity storage; metrics=latency; throughput; lifetime; results=Latency -8–53%; throughput +49–316%; lifetime +268–327%.
Interpretation: 가용 용량을 런타임 제어 변수로 만들면 성능·endurance를 workload에 맞출 수 있다.
Reusable lesson: 스토리지 용량도 고정 자원이 아니라 성능·수명용 탄력 자원으로 보라.
Applicability: flash SSD와 로그 구조 파일시스템.
Limits: 사용자 용량을 희생하며 custom SSD·FS·manager와 수명 가정이 필요하다. +
Verification: abstract_only; confidence=high.
Canonical title: The Design and Implementation of a Capacity-Variant Storage System
Question: SSD가 수명·성능을 위해 실행 중 용량을 탄력적으로 줄일 수 있는가?
Context: 고정 over-provisioning은 workload 변화에 따라 용량·GC·수명 절충을 최적화하지 못한다.
Method: CV-SSD, 탄력 로그 구조 CV-FS, CV-manager가 용량을 점진적으로 조정한다.
Evaluation: workloads=synthetic and real workloads; baselines=fixed-capacity storage; metrics=latency; throughput; lifetime; results=Latency -8–53%; throughput +49–316%; lifetime +268–327%.
Interpretation: 가용 용량을 런타임 제어 변수로 만들면 성능·endurance를 workload에 맞출 수 있다.
Reusable lesson: 스토리지 용량도 고정 자원이 아니라 성능·수명용 탄력 자원으로 보라.
Applicability: flash SSD와 로그 구조 파일시스템.
Limits: 사용자 용량을 희생하며 custom SSD·FS·manager와 수명 가정이 필요하다. +
Bibliographic paper record. +
Bibliographic paper record. +
Verification: full_text; confidence=high.
Canonical title: Tiered Memory Management: Access Latency is the Key!
Question: Should tiered-memory placement optimize access frequency or the loaded latency each tier experiences under contention?
Context: Hot-page packing assumes fixed tier speed, but queueing can make a nominally fast tier slower as load rises.
Method: Colloid estimates each tier's loaded latency from hardware counters and Little's Law, then moves pages to balance average access latency; it integrates with HeMem, TPP, and MEMTIS.
Evaluation: workloads=GUPS; real applications; static and time-varying loads; baselines=HeMem; TPP; MEMTIS; offline optimal; metrics=loaded access latency; application performance; distance from optimal; results=loaded latency up to 5x unloaded; prior managers 2.30–2.46x worse than optimal; Colloid near optimal
Interpretation: The right placement objective is marginal loaded latency, not static media class or hotness alone.
Reusable lesson: Control placement with measured queueing-sensitive latency and balance load across tiers.
Applicability: NUMA-, CXL-, and heterogeneous-memory tiering with separate channels.
Limits: Requires distinguishable tier traffic and reliable counters; evaluation uses NUMA as a tiered-memory proxy. +
Verification: full_text; confidence=high.
Canonical title: Tiered Memory Management: Access Latency is the Key!
Question: Should tiered-memory placement optimize access frequency or the loaded latency each tier experiences under contention?
Context: Hot-page packing assumes fixed tier speed, but queueing can make a nominally fast tier slower as load rises.
Method: Colloid estimates each tier's loaded latency from hardware counters and Little's Law, then moves pages to balance average access latency; it integrates with HeMem, TPP, and MEMTIS.
Evaluation: workloads=GUPS; real applications; static and time-varying loads; baselines=HeMem; TPP; MEMTIS; offline optimal; metrics=loaded access latency; application performance; distance from optimal; results=loaded latency up to 5x unloaded; prior managers 2.30–2.46x worse than optimal; Colloid near optimal
Interpretation: The right placement objective is marginal loaded latency, not static media class or hotness alone.
Reusable lesson: Control placement with measured queueing-sensitive latency and balance load across tiers.
Applicability: NUMA-, CXL-, and heterogeneous-memory tiering with separate channels.
Limits: Requires distinguishable tier traffic and reliable counters; evaluation uses NUMA as a tiered-memory proxy. +
Verification: official USENIX page and abstract; confidence=high.
Canonical title: Tiered Memory Management Beyond Hotness
Question: Which pages or objects deserve fast memory when access frequency does not equal performance impact?
Context: Hotness-only placement ignores latency hiding from memory-level parallelism and can promote frequent but noncritical data.
Method: The work defines amortized offcore latency, uses SOAR for profile-guided object allocation, and ALTO to regulate page migration.
Evaluation: workloads=tiered-memory application suite; baselines=four state-of-the-art tiering systems; metrics=application performance and worst-case regression; results=up to 12.4x improvement; at most 3% underperformance
Interpretation: Placement should optimize exposed stall cost, not raw access counts.
Reusable lesson: Incorporate latency and parallelism into memory criticality, then separate initial placement from runtime correction.
Applicability: DRAM plus slower NUMA/CXL/persistent-memory tiers.
Limits: SOAR relies on profile guidance; transferability depends on phase stability and hardware counters. +
Bibliographic paper record. +
Verification: official USENIX page and abstract; confidence=medium.
Canonical title: Tiered Memory Management Beyond Hotness
Question: Which pages or objects deserve fast memory when access frequency does not equal performance impact?
Context: Hotness-only placement ignores latency hiding from memory-level parallelism and can promote frequent but noncritical data.
Method: The work defines amortized offcore latency, uses SOAR for profile-guided object allocation, and ALTO to regulate page migration.
Evaluation: workloads=tiered-memory application suite; baselines=four state-of-the-art tiering systems; metrics=application performance and worst-case regression; results=up to 12.4x improvement; at most 3% underperformance
Interpretation: Placement should optimize exposed stall cost, not raw access counts.
Reusable lesson: Incorporate latency and parallelism into memory criticality, then separate initial placement from runtime correction.
Applicability: DRAM plus slower NUMA/CXL/persistent-memory tiers.
Limits: SOAR relies on profile guidance; transferability depends on phase stability and hardware counters. +
Bibliographic paper record. +
Verification: official DOI/EuroSys metadata plus primary Intel artifact and Zenodo record; confidence=high.
Canonical title: TierScape: Harnessing Multiple Compressed Tiers to Tame Server Memory TCO
Question: Can multiple byte-addressable and compressed memory tiers reduce server-memory TCO while preserving performance?
Context: A single DRAM/slow-memory or one compressed tier exposes a coarse cost/performance tradeoff.
Method: TierScape supports multiple byte-addressable and compressible tiers and compares baseline, HeMem, ILP, and Waterfall placement strategies using PEBS hotness.
Evaluation: workloads=MASIM and 40 GB Memcached with 4 KB objects; baselines=no tiering, HeMem, ILP, Waterfall; metrics=performance, tier distribution, pressure stalls, memory TCO; results=The primary artifact evaluates MASIM and a 40 GB Memcached dataset with 4 KB objects on DRAM/Optane; the accessible primary pages did not state an unambiguous headline speedup/TCO number.
Interpretation: Multiple configurable tiers offer a richer placement frontier than a fixed two-tier design.
Reusable lesson: Model placement as a cost/performance optimization across several tier types, not a binary hot/cold split.
Applicability: NUMA/CXL/Optane servers with byte-addressable and optional compressed tiers.
Limits: Reproducing performance requires Optane-like latency; compression needs a Linux 5.17 patch, at least two NUMA nodes, and artifact result reproduction is marked TODO. +
Verification: official DOI/EuroSys metadata plus primary Intel artifact and Zenodo record; confidence=medium.
Canonical title: TierScape: Harnessing Multiple Compressed Tiers to Tame Server Memory TCO
Question: Can multiple byte-addressable and compressed memory tiers reduce server-memory TCO while preserving performance?
Context: A single DRAM/slow-memory or one compressed tier exposes a coarse cost/performance tradeoff.
Method: TierScape supports multiple byte-addressable and compressible tiers and compares baseline, HeMem, ILP, and Waterfall placement strategies using PEBS hotness.
Evaluation: workloads=MASIM and 40 GB Memcached with 4 KB objects; baselines=no tiering, HeMem, ILP, Waterfall; metrics=performance, tier distribution, pressure stalls, memory TCO; results=The primary artifact evaluates MASIM and a 40 GB Memcached dataset with 4 KB objects on DRAM/Optane; the accessible primary pages did not state an unambiguous headline speedup/TCO number.
Interpretation: Multiple configurable tiers offer a richer placement frontier than a fixed two-tier design.
Reusable lesson: Model placement as a cost/performance optimization across several tier types, not a binary hot/cold split.
Applicability: NUMA/CXL/Optane servers with byte-addressable and optional compressed tiers.
Limits: Reproducing performance requires Optane-like latency; compression needs a Linux 5.17 patch, at least two NUMA nodes, and artifact result reproduction is marked TODO. +
Verification: arXiv paper/abstract and official DOI metadata; confidence=high.
Canonical title: FDP: Flexible Data Placement for Flash Caches
Question: Can flash caches reduce internal garbage collection and write amplification by exposing data lifetime?
Context: Mixed-lifetime cache objects cause NAND garbage collection to copy still-live data during reclamation.
Method: The work integrates NVMe Flexible Data Placement into CacheLib and segregates writes by expected lifetime.
Evaluation: workloads=Meta/Twitter production cache traces; baselines=conventional flash-cache placement; metrics=device write amplification, garbage collection, overhead; results=write amplification approximately 1 with little overhead
Interpretation: Host-known object lifetime is sufficient to make flash placement substantially more efficient.
Reusable lesson: Pass lifecycle hints across the host/device boundary instead of forcing the FTL to infer them.
Applicability: Large flash caches on FDP-capable SSDs with classifiable object lifetimes.
Limits: Requires FDP support and useful lifetime classification; results reflect production cache configurations. +