속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
d
Verification: arXiv full-text excerpts, DOI metadata, and official acceptance page; confidence=high.
Canonical title: DeltaZip: Multi-Tenant Language Model Serving via Delta Compression
Question: How can many fine-tuned LLM variants be served without storing and loading a full weight copy per tenant?
Context: Fine-tuned models share a base lineage but conventional runtimes duplicate weights and serialize costly model swaps.
Method: DeltaZip stores a shared base plus structured-sparse, quantized, optionally lossless deltas and executes custom sparse-batched matrix multiplication in a vLLM-based server.
Evaluation: workloads=fine-tuned Llama-family variants and multi-tenant request traces; baselines=vLLM and conventional model swapping; metrics=delta size, throughput, end-to-end latency, TTFT; results=up to 13x delta compression; 2–12x throughput; 1.6–16x latency/TTFT improvements
Interpretation: Cross-model weight similarity is a serving resource that can be exploited without restricting adaptation to PEFT.
Reusable lesson: Represent related model variants as compact executable deltas, not independent checkpoints.
Applicability: Multi-tenant serving of full- or parameter-efficiently fine-tuned models sharing a base.
Limits: Prompt computation is not accelerated; gains narrow at uniformly high load, and compression needs calibration plus known base lineage. +
Source: Yonsei University Computer Systems Laboratory publication list supplied by the user. Manifestation 2 of 2. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_demand_based_coordinated_scheduling_for_smp_vms_3d4e78ab. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1145/2451116.2451156 ; https://doi.org/10.1145/2499368.2451156 ; https://yonsei.elsevierpure.com/en/publications/demand-based-coordinated-scheduling-for-smp-vms/. 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 2. +
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_development_of_behavior_profilers_for_multimedia_consumer_electronics_f30d374c. 확인 범위: official_abstract. 확인한 자료: https://doi.org/10.1109/TCE.2009.5373752 ; https://yonsei.elsevierpure.com/en/publications/development-of-behavior-profilers-for-multimedia-consumer-electro/. 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_dflop_a_data_driven_framework_for_multimodal_llm_training_pipeline_optimization_9bf046db. 확인 범위: full_text. 확인한 자료: https://arxiv.org/abs/2603.25120 ; https://arxiv.org/pdf/2603.25120 ; https://www.microsoft.com/en-us/research/publication/dflop-a-data-driven-framework-for-multimodal-llm-training-pipeline-optimization/. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Bibliographic paper record. +
Verification: official_abstract; confidence=high.
Canonical title: Disaggregated RAID Storage in Modern Datacenters
Question: How should RAID be redesigned when disks, compute, and network are disaggregated across a datacenter?
Context: Centralized RAID controllers and blocking multistage writes underuse peer bandwidth and magnify network overhead.
Method: dRAID enables peer-to-peer disaggregated data access, nonblocking multistage writes, pipelined I/O, and bandwidth-aware reconstruction.
Evaluation: workloads=dRAID microbenchmarks; object-store workloads; baselines=conventional disaggregated RAID; metrics=bandwidth; object-store throughput; reconstruction efficiency; results=up to 3x bandwidth; 1.5–2.35x object-store throughput
Interpretation: RAID control/data paths should exploit the network topology and overlap protection work rather than emulate a local controller.
Reusable lesson: Redesign recovery and parity flows around disaggregated peer bandwidth, not a centralized legacy abstraction.
Applicability: Datacenter disaggregated storage and object stores.
Limits: Benefits depend on topology, network balance, failure model, reconstruction traffic, and object-store integration. +
Verification: official_abstract; confidence=medium.
Canonical title: Disaggregated RAID Storage in Modern Datacenters
Question: How should RAID be redesigned when disks, compute, and network are disaggregated across a datacenter?
Context: Centralized RAID controllers and blocking multistage writes underuse peer bandwidth and magnify network overhead.
Method: dRAID enables peer-to-peer disaggregated data access, nonblocking multistage writes, pipelined I/O, and bandwidth-aware reconstruction.
Evaluation: workloads=dRAID microbenchmarks; object-store workloads; baselines=conventional disaggregated RAID; metrics=bandwidth; object-store throughput; reconstruction efficiency; results=up to 3x bandwidth; 1.5–2.35x object-store throughput
Interpretation: RAID control/data paths should exploit the network topology and overlap protection work rather than emulate a local controller.
Reusable lesson: Redesign recovery and parity flows around disaggregated peer bandwidth, not a centralized legacy abstraction.
Applicability: Datacenter disaggregated storage and object stores.
Limits: Benefits depend on topology, network balance, failure model, reconstruction traffic, and object-store integration. +
Verification: full_text; confidence=high.
Question: How can one service many LoRA adapters while adapting to changing request skew and GPU load?
Context: Always merging adapters is slow to switch; never merging them loses kernel efficiency and can create imbalance.
Method: dLoRA uses credits to decide adapter merge/unmerge and jointly migrates requests with adapters across replicas.
Evaluation: workloads=multi-adapter LoRA serving workloads; baselines=vLLM; Hugging Face PEFT; S-LoRA; metrics=throughput; average latency; results=up to 57.9x throughput vs vLLM; up to 26x vs PEFT; up to 1.8x lower latency vs S-LoRA
Interpretation: Adapter representation and request placement must be orchestrated together at runtime.
Reusable lesson: Dynamically switch between general and specialized execution paths using observed demand and switching cost.
Applicability: Multi-tenant LoRA serving with many adapters and skewed arrivals.
Limits: Performance depends on adapter popularity, merge/migration cost, model size, and workload dynamics. +
Verification: full_text; confidence=high.
Question: How can one service many LoRA adapters while adapting to changing request skew and GPU load?
Context: Always merging adapters is slow to switch; never merging them loses kernel efficiency and can create imbalance.
Method: dLoRA uses credits to decide adapter merge/unmerge and jointly migrates requests with adapters across replicas.
Evaluation: workloads=multi-adapter LoRA serving workloads; baselines=vLLM; Hugging Face PEFT; S-LoRA; metrics=throughput; average latency; results=up to 57.9x throughput vs vLLM; up to 26x vs PEFT; up to 1.8x lower latency vs S-LoRA
Interpretation: Adapter representation and request placement must be orchestrated together at runtime.
Reusable lesson: Dynamically switch between general and specialized execution paths using observed demand and switching cost.
Applicability: Multi-tenant LoRA serving with many adapters and skewed arrivals.
Limits: Performance depends on adapter popularity, merge/migration cost, model size, and workload dynamics. +
Bibliographic paper record. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_dram_0a5cae63. 확인 범위: official_abstract. 확인한 자료: https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART003203192. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility. +
Full Korean bibliographic citation is preserved in the lesson evidence field; this ASCII citation is used for adapter compatibility. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_dram_architecture_for_efficient_data_lifetime_management_0f2dfdbe. 확인 범위: full_text. 확인한 자료: 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. 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. +
legacy tier mode와 corrected independent mode의 설계 및 결과 반전을 확인함. +