PoC placement_id 결정 코드 path

세부 분석은 01_codebase_map.md §4 와 03_fdp_implementation_audit.md (a) 에 정리. 본 노트는 핵심 함수 한 곳에 모아둔 빠른 레퍼런스.

Data 경로

# lmcache/v1/storage_backend/raw_block/core.py:342
def _select_fdp_ruh(self, key: RawBlockKeySpec, obj: MemoryObj | None = None) -> int | None:
    if not self.use_fdp:
        return None

    placement_rank = (
        getattr(obj.metadata, "fdp_placement_rank", None)
        if obj is not None else None
    )
    if placement_rank is not None:
        worker_index = int(placement_rank)
    elif self.key_namespace == "object":
        kv_rank = int(decode_object_key(key.encoded).kv_rank)
        # ObjectKey.ComputeKVRank packs global_rank into bits [16, 23].
        worker_index = (kv_rank >> 16) & 0xFF
    else:
        parsed_key = decode_legacy_key(key.encoded)
        worker_index = int(getattr(parsed_key, "worker_id", 0))

    idx = worker_index % len(self.fdp_data_ruh_ids)
    return self.fdp_data_ruh_ids[idx]

3단 fallback:

1. obj.metadata.fdp_placement_rank — MP worker가 stamp (lmcache/v1/multiprocess/server.py:34)
2. object key의 kv_rank>>16 — global rank의 worker index 비트
3. legacy key의 worker_id 필드

→ worker_index % len(fdp_data_ruh_ids) 로 RUH index 결정.

Metadata 경로

# lmcache/v1/storage_backend/raw_block/core.py:380
def _select_fdp_metadata_ruh(self, encoded_key: str) -> int | None:
    if not self.use_fdp:
        return None
    digest = hashlib.blake2b(encoded_key.encode("utf-8"), digest_size=8).digest()
    idx = int.from_bytes(digest, "little", signed=False) % len(self.fdp_metadata_ruh_ids)
    return self.fdp_metadata_ruh_ids[idx]

→ blake2b 해시 기반 분배. Worker 분배와 독립. metadata와 data는 RUH pool도 별개 (fdp_data_ruh_ids vs fdp_metadata_ruh_ids).

Directive 산출

# lmcache/v1/storage_backend/raw_block/core.py:399
def _write_directive_for_ruh(self, fdp_ruh_id: int | None) -> tuple[int, int]:
    if fdp_ruh_id is None or not self.use_fdp:
        return (0, 0)
    return (self.fdp_directive_type, int(fdp_ruh_id))   # 기본 (2, ruh_id)

Rust 측

// rust/raw_block/src/lib.rs (PoC HEAD)
fn validate_nvme_directive(use_uring_cmd: bool, dtype: u8, dspec: u16) -> PyResult<()> {
    if dtype > 0x0f { return Err(...); }
    if dspec != 0 && dtype == 0 { return Err(...); }
    if (dtype != 0 || dspec != 0) && !use_uring_cmd { return Err(...); }
    Ok(())
}

// nvme_uring_cmd_prep 안:
// cmd.cdw13 = (dspec as u32) << 16     ← placement ID 위치 (cdw13 bits 16-31)
// cmd.cdw12 |= (dtype as u32) & 0x0F   ← directive type (cdw12 bits 0-3)

정책 외부화 (harness layer)

코어 레이어는 worker 단위 분배만 제공. prompt-aware / phase-aware 정책은 harness가 RUH ID 자체를 직접 결정해서 config로 내려주는 방식:

# benchmarks/agentic_mp_trace/replay/fdp_policy.py:18
def resolve_policy(*, mode_cfg, storage_class, ruh_count) -> tuple[bool, list[int], list[int]]:
    use_fdp = bool(mode_cfg.get("use_fdp", False))
    if not use_fdp: return False, [], []
    data_spec = mode_cfg.get("default_data_ruhs")
    metadata_spec = mode_cfg.get("default_metadata_ruhs")
    class_cfg = mode_cfg.get("classes", {}).get(storage_class)
    if class_cfg is not None:
        data_spec = class_cfg.get("data_ruhs", data_spec)
        metadata_spec = class_cfg.get("metadata_ruhs", metadata_spec)
    return True, expand_ruh_ids(data_spec), expand_ruh_ids(metadata_spec)

YAML config 예 (benchmarks/fdp_waf_stress/config.example.yaml):

modes:
  separated:
    use_fdp: true
    default_data_ruhs: "0-2"
    default_metadata_ruhs: [3]
    classes:
      hot_prompt:
        data_ruhs: [0]
      cold_prompt:
        data_ruhs: [1, 2]

→ harness가 storage_class 단위로 RUH list를 결정해서 LMCache adapter config로 전달. 코어는 그저 list에서 worker_index % N로 분배.