| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """P2PSecondaryTierManager: Secondary tier for P2P KV cache sharing. |
| 4 | |
| 5 | Owns transports and a single bidirectional P2PSession per remote peer. |
| 6 | """ |
| 7 | |
| 8 | from __future__ import annotations |
| 9 | |
| 10 | import time |
| 11 | import uuid |
| 12 | from collections.abc import Iterable, Sequence |
| 13 | from dataclasses import dataclass, field |
| 14 | from typing import TYPE_CHECKING, Any |
| 15 | |
| 16 | from typing_extensions import override |
| 17 | |
| 18 | import vllm.envs as envs |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.v1.core.kv_cache_utils import get_none_hash_seed |
| 21 | from vllm.v1.kv_offload.base import ( |
| 22 | LookupResult, |
| 23 | OffloadKey, |
| 24 | OffloadPolicy, |
| 25 | ReqContext, |
| 26 | RequestOffloadingContext, |
| 27 | ScheduleEndContext, |
| 28 | ) |
| 29 | from vllm.v1.kv_offload.file_mapper import FileMapper |
| 30 | from vllm.v1.kv_offload.tiering.base import ( |
| 31 | JobResult, |
| 32 | SecondaryTierManager, |
| 33 | TransferJob, |
| 34 | ) |
| 35 | from vllm.v1.kv_offload.tiering.p2p.control import ControlTransport, ZmqTransport |
| 36 | from vllm.v1.kv_offload.tiering.p2p.data import DataTransport, NixlTransport |
| 37 | from vllm.v1.kv_offload.tiering.p2p.session import P2PSession |
| 38 | |
| 39 | if TYPE_CHECKING: |
| 40 | from vllm.v1.kv_offload.base import OffloadingSpec |
| 41 | from vllm.v1.kv_offload.tiering.base import ParentManager |
| 42 | from vllm.v1.kv_offload.tiering.p2p.control.base import ControlConnection |
| 43 | |
| 44 | logger = init_logger(__name__) |
| 45 | |
| 46 | # Reap unbound store batches that have been parked without a FetchMsg |
| 47 | # binding them to a session for longer than this. Protects against the |
| 48 | # prefiller buffering blocks for a decoder that never asks (decoder died, |
| 49 | # network partition, lost kv_request_id). Applies only while a batch is |
| 50 | # unbound: binding hands it to the session, which times the job out under |
| 51 | # _STORE_TIMEOUT_S instead. |
| 52 | # Overridable per tier via the ``unbound_store_timeout_s`` config key. |
| 53 | _UNBOUND_STORE_TIMEOUT_S = 60.0 |
| 54 | |
| 55 | # How long a reaped kv_request_id is remembered so a fetch that arrives |
| 56 | # after the reap can be rejected immediately. Must exceed the consumer's |
| 57 | # load timeout (session.client._LOAD_TIMEOUT_S): past that the consumer has |
| 58 | # already given up on its own, so there is nobody left to reject. |
| 59 | _REAPED_ID_RETENTION_S = 60.0 |
| 60 | |
| 61 | # Time we wait during shutdown for inflight transfers to drain via |
| 62 | # cancel(mode="wait") before falling back to mode="immediate". Bounded |
| 63 | # so a wedged peer can't hang shutdown. |
| 64 | _SHUTDOWN_DRAIN_TIMEOUT_S = 3.0 |
| 65 | |
| 66 | # Sleep between iterations of the bounded drain loops in drain_jobs() and |
| 67 | # _drain_inflight_for_shutdown(). Short enough to keep latency low, long |
| 68 | # enough to avoid busy-spinning the scheduler thread. |
| 69 | _DRAIN_SLEEP_S = 0.001 |
| 70 | |
| 71 | |
| 72 | def _remote_prefiller_params(kv_params: dict | None) -> dict | None: |
| 73 | """Return the ``remote_prefiller`` sub-dict, or None if absent. |
| 74 | |
| 75 | Set on decoder requests to name the remote prefiller they pull from; |
| 76 | carries kv_request_id, remote_host, remote_port. |
| 77 | """ |
| 78 | if not kv_params: |
| 79 | return None |
| 80 | return kv_params.get("remote_prefiller") |
| 81 | |
| 82 | |
| 83 | def _remote_decoder_params(kv_params: dict | None) -> dict | None: |
| 84 | """Return the ``remote_decoder`` sub-dict, or None if absent. |
| 85 | |
| 86 | Set on prefiller requests to name the remote decoder they serve; |
| 87 | carries kv_request_id. |
| 88 | """ |
| 89 | if not kv_params: |
| 90 | return None |
| 91 | return kv_params.get("remote_decoder") |
| 92 | |
| 93 | |
| 94 | def _remote_kv_source_params(kv_params: dict | None) -> dict | None: |
| 95 | """Return the ``remote_kv_source`` sub-dict, or None if absent. |
| 96 | |
| 97 | Set on symmetric-P2P consumer requests to name the remote source they |
| 98 | pull from; carries kv_request_id, remote_host, remote_port. |
| 99 | """ |
| 100 | if not kv_params: |
| 101 | return None |
| 102 | return kv_params.get("remote_kv_source") |
| 103 | |
| 104 | |
| 105 | def _peer_id_from_params(role_params: dict) -> str | None: |
| 106 | """Build ``host:port`` peer_id from a role-scoped sub-dict, or None.""" |
| 107 | host = role_params.get("remote_host") |
| 108 | port = role_params.get("remote_port") |
| 109 | if host and port: |
| 110 | return f"{host}:{port}" |
| 111 | return None |
| 112 | |
| 113 | |
| 114 | @dataclass(slots=True) |
| 115 | class P2PSourceInfo: |
| 116 | """Consumer side: this request fetches from a remote (prefiller or peer).""" |
| 117 | |
| 118 | kv_request_id: str |
| 119 | peer_id: str |
| 120 | do_probe: bool # False for remote_prefiller (PD), True for remote_kv_source |
| 121 | |
| 122 | |
| 123 | @dataclass(slots=True) |
| 124 | class P2PDestInfo: |
| 125 | """Producer side: a remote fetches this request's blocks from us. |
| 126 | |
| 127 | ``kv_request_id`` is None when the ``remote_decoder`` block is present |
| 128 | but malformed (no id); the block's presence still marks the request as |
| 129 | remote-decode, so submit_store must fail rather than store locally. |
| 130 | """ |
| 131 | |
| 132 | kv_request_id: str | None |
| 133 | |
| 134 | |
| 135 | def _parse_source(kv_params: dict | None) -> P2PSourceInfo | None: |
| 136 | """Parse the consumer sub-dict (PD ``remote_prefiller`` or symmetric |
| 137 | ``remote_kv_source``) into a ``P2PSourceInfo``, or None if absent/incomplete.""" |
| 138 | role = _remote_prefiller_params(kv_params) |
| 139 | do_probe = False |
| 140 | if role is None: |
| 141 | role = _remote_kv_source_params(kv_params) |
| 142 | do_probe = True |
| 143 | if not role: |
| 144 | return None |
| 145 | peer_id = _peer_id_from_params(role) |
| 146 | kv_request_id = role.get("kv_request_id") |
| 147 | if peer_id is None or not kv_request_id: |
| 148 | return None |
| 149 | return P2PSourceInfo( |
| 150 | kv_request_id=kv_request_id, |
| 151 | peer_id=peer_id, |
| 152 | do_probe=do_probe, |
| 153 | ) |
| 154 | |
| 155 | |
| 156 | def _parse_dest(kv_params: dict | None) -> P2PDestInfo | None: |
| 157 | """Parse the producer ``remote_decoder`` sub-dict into a ``P2PDestInfo``, |
| 158 | or None if the block is absent (not a remote-decode request).""" |
| 159 | role = _remote_decoder_params(kv_params) |
| 160 | if role is None: |
| 161 | return None |
| 162 | return P2PDestInfo(kv_request_id=role.get("kv_request_id") or None) |
| 163 | |
| 164 | |
| 165 | def _annotate_req_context(req_context: ReqContext) -> None: |
| 166 | """Parse kv_transfer_params once and cache the P2P routing state. |
| 167 | |
| 168 | Called from ``on_new_request``; later calls for the same request read |
| 169 | the cached ``P2PSourceInfo``/``P2PDestInfo`` via ``get_state`` instead |
| 170 | of re-parsing. |
| 171 | """ |
| 172 | source = _parse_source(req_context.kv_transfer_params) |
| 173 | if source is not None: |
| 174 | req_context.set_state(source) |
| 175 | dest = _parse_dest(req_context.kv_transfer_params) |
| 176 | if dest is not None: |
| 177 | req_context.set_state(dest) |
| 178 | |
| 179 | |
| 180 | @dataclass |
| 181 | class _UnboundStoreBatch: |
| 182 | """A submit_store batch parked at the manager before any peer has fetched. |
| 183 | |
| 184 | Indexed by kv_request_id only — the prefiller no longer learns the peer |
| 185 | identity at store time. When a FetchMsg(kv_request_id) arrives on some |
| 186 | session, the manager binds the kv_request_id to that session and replays |
| 187 | every parked batch into ServerRole via session.add_stored_blocks. |
| 188 | """ |
| 189 | |
| 190 | job_id: int |
| 191 | keys: list[OffloadKey] |
| 192 | block_ids: Sequence[int] |
| 193 | submitted_at: float = field(default_factory=time.monotonic) |
| 194 | |
| 195 | |
| 196 | class P2PSecondaryTierManager(SecondaryTierManager): |
| 197 | """Secondary tier for P2P KV cache sharing. |
| 198 | |
| 199 | A single P2PSession per remote peer handles both client-role (loading |
| 200 | blocks from the peer) and server-role (serving blocks to the peer) |
| 201 | over the same control connection. |
| 202 | |
| 203 | Single-threaded: every public method runs on the scheduler thread, and |
| 204 | the engine drives polling via ``get_finished_jobs()`` once per step. |
| 205 | ``has_pending_work()`` keeps the engine ticking so the control transport |
| 206 | and existing sessions are polled even when no requests are scheduled. |
| 207 | """ |
| 208 | |
| 209 | def __init__( |
| 210 | self, |
| 211 | offloading_spec: OffloadingSpec, |
| 212 | primary_kv_view: memoryview, |
| 213 | tier_type: str = "p2p", |
| 214 | host: str | None = None, |
| 215 | port: int | None = None, |
| 216 | backends: list[str] | None = None, |
| 217 | num_threads: int = 4, |
| 218 | unbound_store_timeout_s: float = _UNBOUND_STORE_TIMEOUT_S, |
| 219 | **kwargs: Any, |
| 220 | ) -> None: |
| 221 | """Initialize the P2P secondary tier manager. |
| 222 | |
| 223 | All keyword arguments after ``primary_kv_view`` come from the |
| 224 | ``secondary_tiers`` entry in ``kv_connector_extra_config``. See |
| 225 | ``docs/features/kv_offloading_usage.md`` for the user-facing |
| 226 | configuration reference. |
| 227 | |
| 228 | Args: |
| 229 | offloading_spec: Owning ``OffloadingSpec`` (provides normalized |
| 230 | model, parallel, and cache layout configuration). |
| 231 | primary_kv_view: Memoryview over the CPU primary tier; the |
| 232 | NIXL agent registers this region for RDMA transfers. |
| 233 | tier_type: Tier identifier (defaults to ``"p2p"``). |
| 234 | host: Address the ZMQ control socket binds to, used verbatim |
| 235 | as both the bind address and the identity peers dial back |
| 236 | (mirrors the NIXL connector's ``VLLM_NIXL_SIDE_CHANNEL_HOST``; |
| 237 | no auto-detection). Defaults to |
| 238 | ``VLLM_P2P_SIDE_CHANNEL_HOST`` (``localhost``) when not set; |
| 239 | must be set to the node's routable IP for cross-host P2P so |
| 240 | remote peers can reach the socket. |
| 241 | port: Base port for the ZMQ control socket. Must be |
| 242 | reachable from peers. Defaults to |
| 243 | ``VLLM_P2P_SIDE_CHANNEL_PORT`` (``5710``) when not set. |
| 244 | The bound port is ``base + data_parallel_index`` so each |
| 245 | DP replica gets a distinct port (one socket per replica, |
| 246 | like NIXL); for DP=1 the offset is 0. |
| 247 | backends: NIXL transport backends (e.g. ``["UCX"]``, |
| 248 | ``["MOONCAKE"]``, ``["LIBFABRIC"]``). Defaults to |
| 249 | ``["UCX"]``. When any non-UCX backend is requested, the |
| 250 | NIXL agent is initialized with ``backends=...``; |
| 251 | otherwise it falls back to a UCX-only agent with |
| 252 | ``num_threads`` threads. |
| 253 | num_threads: NIXL agent worker threads for the UCX-only |
| 254 | branch. Ignored when ``backends`` contains a non-UCX |
| 255 | entry. |
| 256 | unbound_store_timeout_s: Seconds a producer holds stored blocks |
| 257 | for a consumer that has not fetched them yet, before |
| 258 | ``_reap_unbound_stores`` drops them. Raise it for |
| 259 | deployments whose prefills outlast the default; the blocks |
| 260 | keep primary-tier slots pinned for the whole window. |
| 261 | **kwargs: Reserved for future tier-specific options. |
| 262 | |
| 263 | Raises: |
| 264 | ValueError: If ``unbound_store_timeout_s`` is not a positive |
| 265 | number, or anything convertible to one. |
| 266 | |
| 267 | """ |
| 268 | backpressure_detector = kwargs.pop("backpressure_detector", None) |
| 269 | if backpressure_detector is not None: |
| 270 | # The generic (store-latency) detector is unreliable for P2P: in |
| 271 | # PD mode dropping a store can leave the decoder waiting until the |
| 272 | # load timeout instead of failing fast, and rendezvous time makes |
| 273 | # store latency a poor pressure signal. Reject it until a |
| 274 | # P2P-specific fail-fast path exists. |
| 275 | raise ValueError( |
| 276 | "Backpressure is not supported for the P2P secondary tier. " |
| 277 | "The generic store-latency detector cannot fail fast in PD " |
| 278 | "mode and rendezvous time makes store latency an unreliable " |
| 279 | "pressure signal. Remove the 'backpressure' config from the " |
| 280 | "p2p tier." |
| 281 | ) |
| 282 | super().__init__( |
| 283 | offloading_spec, |
| 284 | primary_kv_view, |
| 285 | tier_type, |
| 286 | ) |
| 287 | try: |
| 288 | timeout_s = float(unbound_store_timeout_s) |
| 289 | except (TypeError, ValueError): |
| 290 | timeout_s = float("nan") |
| 291 | if not timeout_s > 0: |
| 292 | raise ValueError( |
| 293 | f"unbound_store_timeout_s must be a positive number, got " |
| 294 | f"{unbound_store_timeout_s!r}" |
| 295 | ) |
| 296 | self._unbound_store_timeout_s = timeout_s |
| 297 | |
| 298 | # Block hashes chain from NONE_HASH (see v1/core/kv_cache_utils.py). |
| 299 | # Peers whose seeds differ compute different hashes for identical |
| 300 | # content, so lookups silently miss and no KV crosses the wire. The |
| 301 | # seed is advertised and verified against each peer during the |
| 302 | # handshake, so a mismatch is rejected loudly instead of degrading |
| 303 | # silently. Resolved lazily in _get_hash_seed: this tier is built |
| 304 | # before init_none_hash runs, and a non-cryptographic hash algorithm |
| 305 | # seeds NONE_HASH randomly, so the value is only known afterwards. |
| 306 | self._hash_seed: str | None = None |
| 307 | if host is None: |
| 308 | host = envs.VLLM_P2P_SIDE_CHANNEL_HOST |
| 309 | if port is None: |
| 310 | port = envs.VLLM_P2P_SIDE_CHANNEL_PORT |
| 311 | # One control socket per DP replica: offset the base by the global |
| 312 | # data-parallel index so replicas on a host don't collide (mirrors |
| 313 | # NIXL). For DP=1 the index is 0, leaving the base port unchanged. |
| 314 | dp_index = offloading_spec.config.parallel.data_parallel_index |
| 315 | port = int(port) + dp_index |
| 316 | # Two decoupled identities: |
| 317 | # _local_id (``host:port``): the ZMQ control identity that peers |
| 318 | # dial back, used verbatim (the socket binds this host/port and |
| 319 | # the address is parsed back into host:port by the remote). |
| 320 | # _nixl_agent_name (uuid4): the NIXL agent name. It is never dialed |
| 321 | # — it travels opaquely inside the agent metadata blob — so it |
| 322 | # only needs to be globally unique. A per-process uuid guarantees |
| 323 | # that even for peers sharing a host:port (mirrors the NIXL |
| 324 | # connector; avoids the "remote agent name equals local" reject). |
| 325 | self._local_id = f"{host}:{port}" |
| 326 | self._nixl_agent_name = str(uuid.uuid4()) |
| 327 | |
| 328 | config_fields = FileMapper.from_offloading_spec( |
| 329 | root_dir="", |
| 330 | offloading_spec=offloading_spec, |
| 331 | blocks_per_file=offloading_spec.blocks_per_chunk, |
| 332 | parallel_agnostic=True, |
| 333 | ).get_run_config() |
| 334 | self._data: DataTransport = NixlTransport( |
| 335 | self._nixl_agent_name, |
| 336 | primary_kv_view, |
| 337 | config_fields=config_fields, |
| 338 | backends=backends, |
| 339 | num_threads=int(num_threads), |
| 340 | ) |
| 341 | self._control: ControlTransport = ZmqTransport(self._local_id, host, port) |
| 342 | |
| 343 | self._sessions: dict[str, P2PSession] = {} |
| 344 | # kv_request_id → session, set when the bound session has received |
| 345 | # FetchMsg for that id. submit_store after binding routes directly |
| 346 | # to the session; before binding, batches are parked in |
| 347 | # _unbound_stores below. Stays in sync with _sessions: entries |
| 348 | # pointing to a reaped session are purged in _reap_dead_sessions. |
| 349 | self._kv_to_session: dict[str, P2PSession] = {} |
| 350 | # kv_request_id → list of batches submit_store'd before any peer |
| 351 | # asked for that id. Drained into a session by _on_session_fetch |
| 352 | # when the corresponding FetchMsg arrives, or surfaced as failures |
| 353 | # by _reap_unbound_stores after _unbound_store_timeout_s. |
| 354 | self._unbound_stores: dict[str, list[_UnboundStoreBatch]] = {} |
| 355 | # kv_request_id → time its parked batches were reaped. Nothing under |
| 356 | # one of these ids can be served any more, so a FetchMsg for it is |
| 357 | # rejected on the spot by _poll_once instead of parking demand until |
| 358 | # the consumer's own load timeout, and submit_store fails rather than |
| 359 | # re-parking under it — a prefill that outran the timeout would |
| 360 | # otherwise keep pinning fresh primary-tier slots, one timeout at a |
| 361 | # time. Entries are pruned after _REAPED_ID_RETENTION_S. |
| 362 | self._reaped_stores: dict[str, float] = {} |
| 363 | |
| 364 | self._finished_jobs: list[JobResult] = [] |
| 365 | # kv_request_ids that hit a transport/session failure; On load lookup() |
| 366 | # rejects them so the request falls back to local prefill. |
| 367 | self._failed_req_ids: set[str] = set() |
| 368 | # Synthetic lookup ctxs from reaped sessions still owing a |
| 369 | # ``parent.on_request_finished`` (the session's failed_serves). The |
| 370 | # dead session had no parent handle at teardown; these are flushed |
| 371 | # at the top of the next ``serve_external_requests`` where the |
| 372 | # handle is valid. |
| 373 | self._failed_serve_ctxs: list[ReqContext] = [] |
| 374 | |
| 375 | # ------------------------------------------------------------------ |
| 376 | # SecondaryTierManager interface |
| 377 | # ------------------------------------------------------------------ |
| 378 | |
| 379 | @override |
| 380 | def lookup(self, key: OffloadKey, req_context: ReqContext) -> LookupResult: |
| 381 | source = req_context.get_state(P2PSourceInfo) |
| 382 | if source is None: |
| 383 | return LookupResult.MISS |
| 384 | if source.kv_request_id in self._failed_req_ids: |
| 385 | return LookupResult.MISS |
| 386 | |
| 387 | # Symmetric-P2P consumer (``remote_kv_source`` sub-dict): probe the |
| 388 | # peer asynchronously. First call registers the (kv_request_id, |
| 389 | # key) entry and returns RETRY; flush_pending_lookups() |
| 390 | # in on_schedule_end batches the LookupMsg; a later step's |
| 391 | # lookup() returns HIT/MISS once LookupRespMsg has arrived. |
| 392 | # PD path (``remote_prefiller`` sub-dict only) keeps the eager HIT. |
| 393 | if source.do_probe: |
| 394 | session = self._sessions.get(source.peer_id) |
| 395 | if session is None: |
| 396 | return LookupResult.MISS |
| 397 | result = session.register_lookup(source.kv_request_id, key) |
| 398 | if result is True: |
| 399 | return LookupResult.HIT |
| 400 | if result is False: |
| 401 | return LookupResult.MISS |
| 402 | return LookupResult.RETRY |
| 403 | |
| 404 | # PD consumer (we are the decoder): all kv blocks should be on the |
| 405 | # prefiller side. Return HIT immediately. |
| 406 | return LookupResult.HIT |
| 407 | |
| 408 | @override |
| 409 | def on_new_request(self, req_context: ReqContext) -> RequestOffloadingContext: |
| 410 | """Parse kv_transfer_params once and open the outbound session. |
| 411 | |
| 412 | Parses the P2P routing state onto ``req_context`` (cached for the |
| 413 | later lookup/submit/finish calls). On the consumer side |
| 414 | (``remote_prefiller`` for PD or ``remote_kv_source`` for symmetric |
| 415 | P2P), open a session toward the producer at remote_host:remote_port |
| 416 | so submit_load can issue FetchMsg as soon as it fires. On the |
| 417 | prefiller side, sessions are created when the consumer's inbound |
| 418 | connection arrives in _accept_new_peers — submit_store no longer |
| 419 | pre-creates anything. |
| 420 | |
| 421 | Producer-leg requests (``remote_decoder`` present) ask for |
| 422 | REQUEST_LEVEL: the peer needs every block of the request, not just |
| 423 | the ones this request computed. |
| 424 | """ |
| 425 | _annotate_req_context(req_context) |
| 426 | source = req_context.get_state(P2PSourceInfo) |
| 427 | if source is not None: |
| 428 | self._get_or_create_session(source.peer_id) |
| 429 | dest = req_context.get_state(P2PDestInfo) |
| 430 | if dest is not None and dest.kv_request_id: |
| 431 | return RequestOffloadingContext(policy=OffloadPolicy.REQUEST_LEVEL) |
| 432 | return RequestOffloadingContext() |
| 433 | |
| 434 | @override |
| 435 | def on_request_finished(self, req_context: ReqContext) -> None: |
| 436 | """Cancels pending loads and prunes session-scoped state. |
| 437 | |
| 438 | Consumer side (``remote_prefiller`` for PD or ``remote_kv_source`` |
| 439 | for symmetric-P2P): looks up the session by peer_id because the |
| 440 | producer's address is what addresses the client-role load to |
| 441 | cancel; also drops any pending symmetric-P2P lookup state via |
| 442 | ``session.finish_request``. |
| 443 | Prefiller side (``remote_decoder`` set): looks up via kv_request_id |
| 444 | because peer_id is no longer carried on store-time |
| 445 | kv_transfer_params; if a session has bound the id, finish it. If |
| 446 | no session has bound the id yet, this is a no-op: parked batches |
| 447 | in `_unbound_stores` are left in place and cleaned up only by |
| 448 | `_reap_unbound_stores` after `_unbound_store_timeout_s`. |
| 449 | """ |
| 450 | source = req_context.get_state(P2PSourceInfo) |
| 451 | dest = req_context.get_state(P2PDestInfo) |
| 452 | kv_request_id = source.kv_request_id if source is not None else None |
| 453 | if kv_request_id is None and dest is not None: |
| 454 | kv_request_id = dest.kv_request_id |
| 455 | if not kv_request_id: |
| 456 | return |
| 457 | self._failed_req_ids.discard(kv_request_id) |
| 458 | |
| 459 | if source is not None: |
| 460 | session = self._sessions.get(source.peer_id) |
| 461 | if session is not None: |
| 462 | session.finish_request(kv_request_id) |
| 463 | return |
| 464 | |
| 465 | # Prefiller-side finish: identify the session via kv_request_id. |
| 466 | session = self._kv_to_session.pop(kv_request_id, None) |
| 467 | if session is not None: |
| 468 | session.finish_request(kv_request_id) |
| 469 | return |
| 470 | |
| 471 | @override |
| 472 | def submit_store(self, job_metadata: TransferJob) -> None: |
| 473 | job_id = job_metadata.job_id |
| 474 | keys = list(job_metadata.keys) |
| 475 | block_ids = job_metadata.chunk_ids.tolist() |
| 476 | |
| 477 | assert len(keys) == len(block_ids) |
| 478 | |
| 479 | dest = job_metadata.req_context.get_state(P2PDestInfo) |
| 480 | logger.debug( |
| 481 | "P2P %s: submit_store ENTRY job_id=%d blocks=%d " |
| 482 | "remote_decoder=%s kv_request_id=%s", |
| 483 | self._local_id, |
| 484 | job_id, |
| 485 | len(block_ids), |
| 486 | dest is not None, |
| 487 | dest.kv_request_id if dest is not None else None, |
| 488 | ) |
| 489 | # Absent ``remote_decoder`` block => not a remote-decode request: |
| 490 | # succeed locally without parking. An empty/malformed dict is still |
| 491 | # a remote-decode signal and must fail the missing-id check below. |
| 492 | if dest is None: |
| 493 | self._finished_jobs.append(JobResult(job_id=job_id, success=True)) |
| 494 | return |
| 495 | |
| 496 | kv_request_id = dest.kv_request_id |
| 497 | if not kv_request_id: |
| 498 | logger.warning( |
| 499 | "P2P %s: submit_store missing kv_request_id", |
| 500 | self._local_id, |
| 501 | ) |
| 502 | self._finished_jobs.append(JobResult(job_id=job_id, success=False)) |
| 503 | return |
| 504 | |
| 505 | # This id's earlier batches were already reaped, so the consumer has |
| 506 | # either given up or will be rejected by the fetch path. Parking more |
| 507 | # would pin primary-tier slots for another full timeout on blocks |
| 508 | # nobody can fetch, so fail the job now — a prefill that outran the |
| 509 | # timeout falls back to local recompute on the decoder either way. |
| 510 | if kv_request_id in self._reaped_stores: |
| 511 | logger.warning( |
| 512 | "P2P %s: submit_store for reaped kv_request_id=%s job_id=%d " |
| 513 | "— failing without parking", |
| 514 | self._local_id, |
| 515 | kv_request_id, |
| 516 | job_id, |
| 517 | ) |
| 518 | self._finished_jobs.append(JobResult(job_id=job_id, success=False)) |
| 519 | return |
| 520 | |
| 521 | # Fast path: a session has already received FetchMsg for this id, |
| 522 | # so we can route the batch straight into its ServerRole. |
| 523 | session = self._kv_to_session.get(kv_request_id) |
| 524 | if session is not None: |
| 525 | session.add_stored_blocks(kv_request_id, keys, block_ids, job_id) |
| 526 | return |
| 527 | |
| 528 | # No session bound yet — park the batch keyed by kv_request_id. |
| 529 | # _on_session_fetch drains it on the first FetchMsg; if no peer |
| 530 | # ever asks, _reap_unbound_stores surfaces the job as failed. |
| 531 | self._unbound_stores.setdefault(kv_request_id, []).append( |
| 532 | _UnboundStoreBatch( |
| 533 | job_id=job_id, |
| 534 | keys=keys, |
| 535 | block_ids=block_ids, |
| 536 | ) |
| 537 | ) |
| 538 | logger.debug( |
| 539 | "P2P %s: parked submit_store kv_request_id=%s job_id=%d blocks=%d", |
| 540 | self._local_id, |
| 541 | kv_request_id, |
| 542 | job_id, |
| 543 | len(block_ids), |
| 544 | ) |
| 545 | |
| 546 | @override |
| 547 | def submit_load(self, job_metadata: TransferJob) -> None: |
| 548 | job_id = job_metadata.job_id |
| 549 | keys = list(job_metadata.keys) |
| 550 | block_ids = job_metadata.chunk_ids |
| 551 | |
| 552 | source = job_metadata.req_context.get_state(P2PSourceInfo) |
| 553 | logger.debug( |
| 554 | "P2P %s: submit_load ENTRY job_id=%d blocks=%d kv_request_id=%s peer=%s", |
| 555 | self._local_id, |
| 556 | job_id, |
| 557 | len(block_ids), |
| 558 | source.kv_request_id if source is not None else None, |
| 559 | source.peer_id if source is not None else None, |
| 560 | ) |
| 561 | if source is None: |
| 562 | logger.debug( |
| 563 | "P2P %s: submit_load job_id=%d FAILED missing consumer params", |
| 564 | self._local_id, |
| 565 | job_id, |
| 566 | ) |
| 567 | self._finished_jobs.append(JobResult(job_id=job_id, success=False)) |
| 568 | return |
| 569 | |
| 570 | kv_request_id = source.kv_request_id |
| 571 | peer_id = source.peer_id |
| 572 | |
| 573 | if not keys: |
| 574 | logger.debug( |
| 575 | "P2P %s: submit_load job_id=%d short-circuit success (no keys)", |
| 576 | self._local_id, |
| 577 | job_id, |
| 578 | ) |
| 579 | self._finished_jobs.append(JobResult(job_id=job_id, success=True)) |
| 580 | return |
| 581 | |
| 582 | session = self._sessions.get(peer_id) |
| 583 | if session is None: |
| 584 | logger.warning( |
| 585 | "P2P %s: submit_load job_id=%d NO SESSION for peer=%s", |
| 586 | self._local_id, |
| 587 | job_id, |
| 588 | peer_id, |
| 589 | ) |
| 590 | self._finished_jobs.append(JobResult(job_id=job_id, success=False)) |
| 591 | self._failed_req_ids.add(kv_request_id) |
| 592 | return |
| 593 | logger.debug( |
| 594 | "P2P %s: submit_load job_id=%d -> request_blocks peer=%s " |
| 595 | "kv_request_id=%s blocks=%d session_ready=%s", |
| 596 | self._local_id, |
| 597 | job_id, |
| 598 | peer_id, |
| 599 | kv_request_id, |
| 600 | len(block_ids), |
| 601 | session.ready, |
| 602 | ) |
| 603 | session.request_blocks(job_id, kv_request_id, keys, block_ids) |
| 604 | |
| 605 | @override |
| 606 | def get_finished_jobs(self) -> Iterable[JobResult]: |
| 607 | # Drive one polling sweep on the scheduler thread, then hand off |
| 608 | # whatever has accumulated. The engine calls this once per step |
| 609 | # (and keeps stepping while has_pending_work() is True). |
| 610 | self._poll_once() |
| 611 | result = self._finished_jobs |
| 612 | self._finished_jobs = [] |
| 613 | return result |
| 614 | |
| 615 | @override |
| 616 | def has_pending_work(self) -> bool: |
| 617 | # The engine tick is the only driver of _control.poll() and |
| 618 | # session.poll(); without it we miss new peer connects and |
| 619 | # inbound fetch messages on existing sessions. Keep the engine |
| 620 | # ticking for the lifetime of this manager. |
| 621 | return True |
| 622 | |
| 623 | @override |
| 624 | def drain_jobs(self) -> None: |
| 625 | """Block until every submitted load/store job has completed or failed. |
| 626 | |
| 627 | Loops calling ``_poll_once()`` until no session has outstanding |
| 628 | inbound loads or in-flight outbound stores. Mid-flight transfers |
| 629 | are NOT cancelled — the caller (``TieringOffloadingManager.reset_cache``) |
| 630 | needs the primary memoryview to be quiescent, not aborted. Results |
| 631 | accumulate in ``_finished_jobs`` and are surfaced by the next |
| 632 | ``get_finished_jobs()`` call. |
| 633 | """ |
| 634 | start = time.monotonic() |
| 635 | warned = False |
| 636 | while True: |
| 637 | self._poll_once() |
| 638 | pending = any(s.has_pending_work for s in self._sessions.values()) |
| 639 | if not pending: |
| 640 | return |
| 641 | if not warned and time.monotonic() - start > 5.0: |
| 642 | logger.warning( |
| 643 | "P2PSecondaryTierManager.drain_jobs: still draining " |
| 644 | "after 5s; a stuck transfer will block the engine.", |
| 645 | ) |
| 646 | warned = True |
| 647 | time.sleep(_DRAIN_SLEEP_S) |
| 648 | |
| 649 | @override |
| 650 | def serve_external_requests(self, parent: ParentManager) -> None: |
| 651 | """Serve inbound peer lookups against the tiering manager. |
| 652 | |
| 653 | Called once per scheduler step (before this tier's |
| 654 | ``on_schedule_end``) with a ``parent`` handle valid only for the |
| 655 | duration of the call — the sole window in which the P2P server |
| 656 | role may query the tiering manager. First release bookkeeping for |
| 657 | the failed serves left by a reaped session, then let every live |
| 658 | session resolve its enqueued inbound LookupMsgs. |
| 659 | """ |
| 660 | if self._failed_serve_ctxs: |
| 661 | for ctx in self._failed_serve_ctxs: |
| 662 | parent.on_request_finished(ctx) |
| 663 | self._failed_serve_ctxs = [] |
| 664 | for session in self._sessions.values(): |
| 665 | session.serve_external_requests(parent) |
| 666 | |
| 667 | @override |
| 668 | def on_schedule_end(self, context: ScheduleEndContext) -> None: |
| 669 | # Flush any p2p lookups aggregated during this step. |
| 670 | # One LookupMsg per (peer, kv_request_id) with unsent entries; |
| 671 | # send-gating happens inside the session if not yet ready. |
| 672 | for session in self._sessions.values(): |
| 673 | session.flush_pending_lookups() |
| 674 | |
| 675 | # ------------------------------------------------------------------ |
| 676 | # Internal |
| 677 | # ------------------------------------------------------------------ |
| 678 | |
| 679 | def _get_hash_seed(self) -> str: |
| 680 | """The seed NONE_HASH was derived from, resolved on first session.""" |
| 681 | if self._hash_seed is None: |
| 682 | self._hash_seed = get_none_hash_seed() |
| 683 | return self._hash_seed |
| 684 | |
| 685 | def _get_or_create_session(self, peer_id: str) -> P2PSession: |
| 686 | """Return the existing session for peer_id, or open one outbound. |
| 687 | |
| 688 | Consumer-side helper for on_new_request: when ``remote_prefiller`` |
| 689 | (PD) or ``remote_kv_source`` (symmetric P2P) is set, the consumer must reach the |
| 690 | producer at peer_id. If we already have a session toward that |
| 691 | peer (from a prior load or a peer-initiated inbound), reuse it; |
| 692 | otherwise open an outbound ControlConnection and build a |
| 693 | connected session. |
| 694 | """ |
| 695 | session = self._sessions.get(peer_id) |
| 696 | if session is not None: |
| 697 | return session |
| 698 | conn = self._control.connect(peer_id) |
| 699 | session = P2PSession( |
| 700 | peer_id=peer_id, |
| 701 | local_id=self._local_id, |
| 702 | transport=self._data, |
| 703 | local_block_len=self._data.block_len, |
| 704 | local_hash_seed=self._get_hash_seed(), |
| 705 | conn=conn, |
| 706 | ) |
| 707 | self._sessions[peer_id] = session |
| 708 | return session |
| 709 | |
| 710 | def _accept_new_peers(self, new_connections: Sequence[ControlConnection]) -> None: |
| 711 | for conn in new_connections: |
| 712 | logger.info( |
| 713 | "P2P %s: accepting incoming connection from %s", |
| 714 | self._local_id, |
| 715 | conn.peer_id, |
| 716 | ) |
| 717 | try: |
| 718 | existing = self._sessions.get(conn.peer_id) |
| 719 | if existing is not None: |
| 720 | raise ValueError(f"duplicate connection from {conn.peer_id}") |
| 721 | self._sessions[conn.peer_id] = P2PSession( |
| 722 | peer_id=conn.peer_id, |
| 723 | local_id=self._local_id, |
| 724 | transport=self._data, |
| 725 | local_block_len=self._data.block_len, |
| 726 | local_hash_seed=self._get_hash_seed(), |
| 727 | conn=conn, |
| 728 | ) |
| 729 | logger.info( |
| 730 | "P2P %s: created connected session for %s", |
| 731 | self._local_id, |
| 732 | conn.peer_id, |
| 733 | ) |
| 734 | except (ValueError, KeyError, TypeError, AssertionError) as exc: |
| 735 | logger.error("P2P %s: rejecting peer: %s", self._local_id, exc) |
| 736 | conn.close() |
| 737 | |
| 738 | def _reap_dead_sessions(self) -> None: |
| 739 | # Reap connected sessions whose connection died — peer is gone. |
| 740 | # Stranded prefiller-side stores are no longer tracked through a |
| 741 | # session (they live in _unbound_stores keyed by kv_request_id); |
| 742 | # _reap_unbound_stores handles their timeout independently. |
| 743 | dead: list[str] | None = None |
| 744 | for pid, s in self._sessions.items(): |
| 745 | if s.connected and not s.alive: |
| 746 | if dead is None: |
| 747 | dead = [] |
| 748 | dead.append(pid) |
| 749 | if dead is None: |
| 750 | return |
| 751 | for pid in dead: |
| 752 | session = self._sessions.pop(pid) |
| 753 | # Purge any kv_request_id → session entries pointing at this |
| 754 | # session so subsequent submit_stores fall back to the unbound |
| 755 | # path (which will time out into failure if no peer rebinds). |
| 756 | stale_kv_ids = [ |
| 757 | kid for kid, s in self._kv_to_session.items() if s is session |
| 758 | ] |
| 759 | for kid in stale_kv_ids: |
| 760 | del self._kv_to_session[kid] |
| 761 | close_result = session.close() |
| 762 | for job_id in close_result.failed_jobs: |
| 763 | self._finished_jobs.append(JobResult(job_id=job_id, success=False)) |
| 764 | for job_id in close_result.failed_stores: |
| 765 | self._finished_jobs.append(JobResult(job_id=job_id, success=False)) |
| 766 | # Fail every client-side request (in-flight loads plus unresolved |
| 767 | # symmetric-P2P probes) toward the dead peer so lookup() returns |
| 768 | # MISS (local prefill) instead of RETRY forever — even if a fresh |
| 769 | # session to the same peer is later opened by another request. |
| 770 | self._failed_req_ids.update(close_result.failed_req_ids) |
| 771 | # Release the TieringManager's per-request bookkeeping for the |
| 772 | # dead session's synthetic lookups on the next serve_external_requests. |
| 773 | self._failed_serve_ctxs.extend(close_result.failed_serves) |
| 774 | self._data.remove_remote_peer(pid) |
| 775 | logger.warning("P2P %s: peer %s down", self._local_id, pid) |
| 776 | |
| 777 | def _reap_unbound_stores(self) -> None: |
| 778 | """Time out submit_store batches that no peer has ever fetched. |
| 779 | |
| 780 | Walks `_unbound_stores` for entries whose oldest batch is older |
| 781 | than `_unbound_store_timeout_s`. Drops the kv_request_id, surfaces |
| 782 | every batched job as failed, and records the id in `_reaped_stores` |
| 783 | so a late inbound FetchMsg is rejected in one round trip and a late |
| 784 | `submit_store` fails instead of re-parking under the reaped id. |
| 785 | |
| 786 | Also expires `_reaped_stores` entries whose consumer can no longer |
| 787 | be waiting, which is why this runs before the empty check below. |
| 788 | """ |
| 789 | now = time.monotonic() |
| 790 | if self._reaped_stores: |
| 791 | retention_deadline = now - _REAPED_ID_RETENTION_S |
| 792 | for kid in [ |
| 793 | kid |
| 794 | for kid, reaped_at in self._reaped_stores.items() |
| 795 | if reaped_at <= retention_deadline |
| 796 | ]: |
| 797 | del self._reaped_stores[kid] |
| 798 | |
| 799 | if not self._unbound_stores: |
| 800 | return |
| 801 | deadline = now - self._unbound_store_timeout_s |
| 802 | expired: list[str] | None = None |
| 803 | for kid, batches in self._unbound_stores.items(): |
| 804 | # Batches are appended in arrival order, so the head is oldest. |
| 805 | if batches and batches[0].submitted_at <= deadline: |
| 806 | if expired is None: |
| 807 | expired = [] |
| 808 | expired.append(kid) |
| 809 | if expired is None: |
| 810 | return |
| 811 | for kid in expired: |
| 812 | batches = self._unbound_stores.pop(kid) |
| 813 | self._reaped_stores[kid] = now |
| 814 | for batch in batches: |
| 815 | self._finished_jobs.append( |
| 816 | JobResult(job_id=batch.job_id, success=False) |
| 817 | ) |
| 818 | logger.warning( |
| 819 | "P2P %s: unbound store kv_request_id=%s timed out after %.0fs " |
| 820 | "without a fetch — failing %d job(s)", |
| 821 | self._local_id, |
| 822 | kid, |
| 823 | self._unbound_store_timeout_s, |
| 824 | len(batches), |
| 825 | ) |
| 826 | |
| 827 | # ------------------------------------------------------------------ |
| 828 | # Polling |
| 829 | # ------------------------------------------------------------------ |
| 830 | |
| 831 | def _poll_once(self) -> None: |
| 832 | """One sweep of the polling work. |
| 833 | |
| 834 | Drains the control transport, polls every session, accumulates |
| 835 | their results into ``_finished_jobs``, and reaps any dead sessions. |
| 836 | Runs on the scheduler thread. |
| 837 | """ |
| 838 | new_connections = self._control.poll() |
| 839 | if new_connections: |
| 840 | logger.info( |
| 841 | "P2P %s: _poll_once got %d new connection(s): %s", |
| 842 | self._local_id, |
| 843 | len(new_connections), |
| 844 | [c.peer_id for c in new_connections], |
| 845 | ) |
| 846 | |
| 847 | self._accept_new_peers(new_connections) |
| 848 | |
| 849 | for session in self._sessions.values(): |
| 850 | result = session.poll() |
| 851 | for lr in result.loads: |
| 852 | self._finished_jobs.append( |
| 853 | JobResult(job_id=lr.job_id, success=lr.success) |
| 854 | ) |
| 855 | if not lr.success: |
| 856 | self._failed_req_ids.add(lr.kv_request_id) |
| 857 | for sr in result.stores: |
| 858 | self._finished_jobs.append( |
| 859 | JobResult(job_id=sr.job_id, success=sr.success) |
| 860 | ) |
| 861 | # Bind kv_request_id → session for any FetchMsg this tick and |
| 862 | # replay any submit_store batches parked while no peer was |
| 863 | # asking. ServerRole.on_fetch already recorded the demand |
| 864 | # inline in dispatch, so the replayed add_stored_blocks calls |
| 865 | # match that demand and submit transfers immediately. |
| 866 | for kv_request_id in result.new_fetch_ids: |
| 867 | if kv_request_id in self._reaped_stores: |
| 868 | # This id's blocks were already reaped, so no |
| 869 | # submit_store will ever satisfy the demand on_fetch |
| 870 | # just recorded. Finalize the round now: the peer gets |
| 871 | # TransferDoneMsg(success=False) on this tick and falls |
| 872 | # back to local prefill, instead of parking until its |
| 873 | # own load timeout expires. The id is deliberately left |
| 874 | # unbound so any later submit_store keeps taking the |
| 875 | # reject path. Batches re-parked between the reap and the |
| 876 | # retention prune are failed here rather than left to a |
| 877 | # second reap, which would pin their slots for another |
| 878 | # full timeout. |
| 879 | session.finish_request(kv_request_id) |
| 880 | for batch in self._unbound_stores.pop(kv_request_id, ()): |
| 881 | self._finished_jobs.append( |
| 882 | JobResult(job_id=batch.job_id, success=False) |
| 883 | ) |
| 884 | continue |
| 885 | self._kv_to_session[kv_request_id] = session |
| 886 | for batch in self._unbound_stores.pop(kv_request_id, ()): |
| 887 | session.add_stored_blocks( |
| 888 | kv_request_id, batch.keys, batch.block_ids, batch.job_id |
| 889 | ) |
| 890 | |
| 891 | self._reap_dead_sessions() |
| 892 | self._reap_unbound_stores() |
| 893 | |
| 894 | # ------------------------------------------------------------------ |
| 895 | # Lifecycle |
| 896 | # ------------------------------------------------------------------ |
| 897 | |
| 898 | @override |
| 899 | def shutdown(self) -> None: |
| 900 | self._drain_inflight_for_shutdown() |
| 901 | for session in self._sessions.values(): |
| 902 | # Orphan ctxs from close() are intentionally dropped: the manager |
| 903 | # is being torn down, so there is no next serve_external_requests |
| 904 | # to flush them and no TieringManager left to release. |
| 905 | session.close() |
| 906 | self._sessions.clear() |
| 907 | self._kv_to_session.clear() |
| 908 | # Surface buffered store jobs as failed so the engine doesn't |
| 909 | # leak them; the manager is going away after this call. |
| 910 | for batches in self._unbound_stores.values(): |
| 911 | for batch in batches: |
| 912 | self._finished_jobs.append( |
| 913 | JobResult(job_id=batch.job_id, success=False) |
| 914 | ) |
| 915 | self._unbound_stores.clear() |
| 916 | self._reaped_stores.clear() |
| 917 | self._control.close() |
| 918 | self._data.close() |
| 919 | |
| 920 | def _drain_inflight_for_shutdown(self) -> None: |
| 921 | """Best-effort drain of inflight transfers before closing _data. |
| 922 | |
| 923 | Mirrors session._drain_abort but as a single bounded loop. Collects |
| 924 | inflight transfer_ids from each session, repeatedly calls |
| 925 | _data.cancel(..., mode="wait") and _data.poll() so handles can |
| 926 | surface as done/failed, and falls back to mode="immediate" once |
| 927 | _SHUTDOWN_DRAIN_TIMEOUT_S elapses so a wedged peer can't hang us. |
| 928 | """ |
| 929 | ids = [tid for s in self._sessions.values() for tid in s._server._inflight] |
| 930 | if not ids: |
| 931 | return |
| 932 | deadline = time.monotonic() + _SHUTDOWN_DRAIN_TIMEOUT_S |
| 933 | still: list[int] = ids |
| 934 | while still and time.monotonic() < deadline: |
| 935 | still = list(self._data.cancel(still, mode="wait")) |
| 936 | if not still: |
| 937 | break |
| 938 | # poll() advances NIXL handle state so the next wait-cancel |
| 939 | # has a chance to release the handles. |
| 940 | self._data.poll() |
| 941 | time.sleep(_DRAIN_SLEEP_S) |
| 942 | if still: |
| 943 | logger.warning( |
| 944 | "P2P %s: shutdown drain timed out after %.1fs with %d " |
| 945 | "transfers still inflight — force-cancelling", |
| 946 | self._local_id, |
| 947 | _SHUTDOWN_DRAIN_TIMEOUT_S, |
| 948 | len(still), |
| 949 | ) |
| 950 | self._data.cancel(still, mode="immediate") |