속성:Reusable lesson
비슷한 문제를 만났을 때 다시 쓸 수 있는 내용을 적습니다.
o
가변 반복 작업은 iteration 경계에서 재배치하라. +
느린 매체를 메모리화할 때 지연 분포와 endurance를 동시에 평가하라. +
ozz identifying kernel out of order concurrency bugs with in vivo memory access reordering 0b85f639 +
Emulate nondeterministic low-level behavior and explore it jointly with software scheduling. +
p
Replicate shared read-heavy kernel data at NUMA boundaries and make ownership transitions explicit. +
Estimate per-item criticality from exposed latency and parallelism, then design migration policies around its skew. +
가속기 스케줄링 정책을 런타임이 명시적으로 제어하게 하라. +
papi exploiting dynamic parallelism in large language model decoding with a processing in memory 5f9f53cc +
Schedule by measured bottleneck class instead of pinning whole models to one device type. +
partial failure resilient memory management system for cxl based distributed shared memory 48948c1d +
분산 메모리 수명 관리는 프로세스 장애 세대를 참조 메타데이터에 넣어라. +
pegasus tolerating skewed workloads in distributed storage with in network coherence directories a4e8c8ec +
핫 키 대응은 선택적 복제와 부하 인지 라우팅을 결합하라. +
GPU 최적화는 연산식만 바꾸기보다 register pressure→occupancy와 다단계 kernel launch/transfer를 함께 측정해 메모리 계층 이동과 kernel fusion을 선택해야 한다. +
평균보다 장치별 고분위 지연과 장기 추세를 감시하라. +
Align placement units with object allocation and locality when pages mix valuable and cold bytes. +
메타데이터 비용이 크면 exact counter를 windowed set membership으로 대체하라. +
pit optimization of dynamic sparse deep learning models via permutation invariant transformation 48509ad0 +
불규칙 데이터는 의미 보존 재배열로 하드웨어 친화적 형태로 변환하라. +
For editable or precompiled ML runtimes, snapshot the entire effective package rather than only Git files or shared objects; reject inherited behavior controls, prove actual import origin, content-bind both package and native subsets, and retain a strict sanitized audit with the result. +
Source-bound experiment lock directories inside the repository must be explicitly ignored; otherwise a post-lock clean-tree check is guaranteed to fail on its own owner metadata. +
For environment-linked native modules, hash logical in-repository module identity plus followed target bytes, exclude absolute target paths from the digest, fail closed on broken or non-regular targets, and recompute the digest at suite completion to close the normal time-of-check/time-of-use gap. +
Complete only the two deterministic target-model generation calibrations on one isolated B200 GPU at a time. Preserve them as structurally promotable, content-addressed migration candidates; never treat B200 latency, throughput, load, policy, or comparative outcomes as final evidence. +
Content-addressed experiment code should hash normalized repository- or package-relative names plus bytes. Test identical trees at different absolute roots and fail closed on filename/content changes or an empty code set. +
For tmux send-keys through SSH, do not assume local one-argument quoting survives. Encode separators explicitly as tmux Space key tokens or preserve quoting inside the remote command string, then immediately verify the pane, lock owner, run directory, and GPU state. +