속성:Evidence note
자료에서 확인할 수 있는 점이나 한계를 적습니다.
w
Verification: metadata_only; confidence=low.
Question:
Context:
Method:
Evaluation: workloads=; baselines=; metrics=; results=
Interpretation:
Reusable lesson:
Applicability: Zoned Namespace SSD parallelism; content not independently verified.
Limits: Only canonical publisher metadata was accessible; method and evaluation are intentionally unresolved. +
Verification: metadata_only; confidence=low.
Question:
Context:
Method:
Evaluation: workloads=; baselines=; metrics=; results=
Interpretation:
Reusable lesson:
Applicability: Zoned Namespace SSD parallelism; content not independently verified.
Limits: Only canonical publisher metadata was accessible; method and evaluation are intentionally unresolved. +
Bibliographic paper record. +
x
Verification: abstract_only; confidence=medium.
Canonical title: XFaaS: Hyperscale and Low Cost Serverless Functions at Meta
Question: 하루 수조 회 함수 호출을 낮은 비용과 안정적 지연으로 처리할 수 있는가?
Context: Meta 내부 FaaS는 10만대 이상 서버에서도 평균 CPU 사용률이 66%에 머물고 locality·burst 문제가 있다.
Method: locality group·warm runtime, 지연 허용 작업 이동, 전역 dispatch, downstream pacing을 결합한다.
Evaluation: workloads=Meta production: trillions of calls/day, >100K servers; baselines=prior Meta FaaS architecture; metrics=CPU utilization; cost; latency; results=Production-scale deployment verified; abstract gives 66% prior average CPU utilization but no single aggregate speedup.
Interpretation: 초대형 FaaS는 단일 스케줄러보다 locality·시간 이동·흐름 제어의 계층적 조합이 필요하다.
Reusable lesson: 함수 배치뿐 아니라 downstream backpressure까지 자원 제어에 포함하라.
Applicability: 하이퍼스케일 사내 서버리스 플랫폼.
Limits: Meta 비공개 환경 중심이고 공개 재현 가능한 baseline·정량 수치가 제한적이다. +
Verification: abstract_only; confidence=high.
Canonical title: XFaaS: Hyperscale and Low Cost Serverless Functions at Meta
Question: 하루 수조 회 함수 호출을 낮은 비용과 안정적 지연으로 처리할 수 있는가?
Context: Meta 내부 FaaS는 10만대 이상 서버에서도 평균 CPU 사용률이 66%에 머물고 locality·burst 문제가 있다.
Method: locality group·warm runtime, 지연 허용 작업 이동, 전역 dispatch, downstream pacing을 결합한다.
Evaluation: workloads=Meta production: trillions of calls/day, >100K servers; baselines=prior Meta FaaS architecture; metrics=CPU utilization; cost; latency; results=Production-scale deployment verified; abstract gives 66% prior average CPU utilization but no single aggregate speedup.
Interpretation: 초대형 FaaS는 단일 스케줄러보다 locality·시간 이동·흐름 제어의 계층적 조합이 필요하다.
Reusable lesson: 함수 배치뿐 아니라 downstream backpressure까지 자원 제어에 포함하라.
Applicability: 하이퍼스케일 사내 서버리스 플랫폼.
Limits: Meta 비공개 환경 중심이고 공개 재현 가능한 baseline·정량 수치가 제한적이다. +
Bibliographic paper record. +
Verification: official USENIX page and abstract; confidence=high.
Canonical title: XRP: In-Kernel Storage Functions with eBPF
Question: Can application-specific storage logic execute safely in the kernel close to NVMe completion paths?
Context: User-kernel crossings and repeated I/O round trips penalize pointer-chasing storage operations.
Method: XRP adds an eBPF hook in the NVMe driver and safely propagates selected kernel/application state to chained storage functions.
Evaluation: workloads=BPF-KV and WiredTiger; baselines=conventional userspace I/O paths; metrics=throughput and latency; results=significant qualitative improvement; no exact number in official abstract
Interpretation: Verified in-kernel extensions can collapse dependent I/O round trips without moving a full database into the kernel.
Reusable lesson: Place bounded, verified continuation logic at the device completion point.
Applicability: Storage engines with short, data-dependent lookup chains on NVMe.
Limits: Logic is constrained by eBPF verification and kernel-state interfaces; exact evaluation details were not extracted. +
Verification: official USENIX page and abstract; confidence=medium.
Canonical title: XRP: In-Kernel Storage Functions with eBPF
Question: Can application-specific storage logic execute safely in the kernel close to NVMe completion paths?
Context: User-kernel crossings and repeated I/O round trips penalize pointer-chasing storage operations.
Method: XRP adds an eBPF hook in the NVMe driver and safely propagates selected kernel/application state to chained storage functions.
Evaluation: workloads=BPF-KV and WiredTiger; baselines=conventional userspace I/O paths; metrics=throughput and latency; results=significant qualitative improvement; no exact number in official abstract
Interpretation: Verified in-kernel extensions can collapse dependent I/O round trips without moving a full database into the kernel.
Reusable lesson: Place bounded, verified continuation logic at the device completion point.
Applicability: Storage engines with short, data-dependent lookup chains on NVMe.
Limits: Logic is constrained by eBPF verification and kernel-state interfaces; exact evaluation details were not extracted. +
Bibliographic paper record. +
z
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_z_journal_scalable_per_core_journaling_1f19a370. 확인 범위: full_text. 확인한 자료: https://www.usenix.org/conference/atc21/presentation/kim-jongseok ; https://www.usenix.org/system/files/atc21-kim-jongseok.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. +
정본 Lesson 보강 근거. 검토 원본: Lesson:technical_review_zero_copying_i_o_stack_for_low_latency_ssds_fbe45009. 확인 범위: official_abstract. 확인한 자료: https://yonsei.elsevierpure.com/en/publications/zero-copying-io-stack-for-low-latency-ssds ; https://doi.org/10.1109/LCA.2021.3064876. 질문, 방법, 평가, 해석, 재사용 교훈, 적용 범위와 한계를 같은 Lesson 본문에 통합했습니다. +