ps/brpc/brpc_ps_client.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "brpc_ps_client.h" | ||
| 2 | |||
| 3 | #include <brpc/channel.h> | ||
| 4 | #include <fmt/format.h> | ||
| 5 | #include <gflags/gflags.h> | ||
| 6 | |||
| 7 | #include <algorithm> | ||
| 8 | #include <cstdint> | ||
| 9 | #include <cstring> | ||
| 10 | #include <future> | ||
| 11 | #include <string> | ||
| 12 | #include <vector> | ||
| 13 | |||
| 14 | #include <google/protobuf/arena.h> | ||
| 15 | |||
| 16 | #include "base/array.h" | ||
| 17 | #include "base/factory.h" | ||
| 18 | #include "base/flatc.h" | ||
| 19 | #include "base/log.h" | ||
| 20 | #include "base/timer.h" | ||
| 21 | #include "ps/base/parameters.h" | ||
| 22 | #include "ps_brpc.pb.h" | ||
| 23 | |||
| 24 | #ifdef ENABLE_PERF_REPORT | ||
| 25 | # include <chrono> | ||
| 26 | # include "base/report/report_client.h" | ||
| 27 | #endif | ||
| 28 | |||
| 29 | using recstoreps_brpc::CommandRequest; | ||
| 30 | using recstoreps_brpc::CommandResponse; | ||
| 31 | using recstoreps_brpc::GetParameterRequest; | ||
| 32 | using recstoreps_brpc::GetParameterResponse; | ||
| 33 | using recstoreps_brpc::InitEmbeddingTableRequest; | ||
| 34 | using recstoreps_brpc::InitEmbeddingTableResponse; | ||
| 35 | using recstoreps_brpc::PSCommand; | ||
| 36 | using recstoreps_brpc::PutParameterRequest; | ||
| 37 | using recstoreps_brpc::PutParameterResponse; | ||
| 38 | using recstoreps_brpc::UpdateParameterRequest; | ||
| 39 | using recstoreps_brpc::UpdateParameterResponse; | ||
| 40 | |||
| 41 | namespace { | ||
| 42 | |||
| 43 | 50 | const ParameterCompressReader* ExtractGetResponseReader( | |
| 44 | const brpc::Controller& cntl, | ||
| 45 | const GetParameterResponse& response, | ||
| 46 | std::string* payload_storage, | ||
| 47 | int* payload_size) { | ||
| 48 | 50 | if (!cntl.response_attachment().empty()) { | |
| 49 | 50 | payload_storage->clear(); | |
| 50 | 50 | cntl.response_attachment().copy_to(payload_storage); | |
| 51 | 50 | *payload_size = payload_storage->size(); | |
| 52 | return reinterpret_cast<const ParameterCompressReader*>( | ||
| 53 | 50 | payload_storage->data()); | |
| 54 | } | ||
| 55 | |||
| 56 | ✗ | *payload_size = response.parameter_value().size(); | |
| 57 | return reinterpret_cast<const ParameterCompressReader*>( | ||
| 58 | ✗ | response.parameter_value().data()); | |
| 59 | } | ||
| 60 | |||
| 61 | } // namespace | ||
| 62 | |||
| 63 | DEFINE_int32(brpc_timeout_ms, 5000, "brpc request timeout in milliseconds"); | ||
| 64 | DEFINE_int32(brpc_max_retry, 3, "brpc max retry times"); | ||
| 65 | DEFINE_bool(parameter_client_random_init_brpc, false, ""); | ||
| 66 | DEFINE_bool(brpc_ps_use_rdma, | ||
| 67 | false, | ||
| 68 | "Use RDMA transport for the brpc PS channel (requires brpc built " | ||
| 69 | "with WITH_RDMA=ON)."); | ||
| 70 | |||
| 71 | namespace { | ||
| 72 | |||
| 73 | // The brpc PS client is loaded as a shared library from Python and never calls | ||
| 74 | // gflags::ParseCommandLineFlags, so RDMA options are configured from env vars | ||
| 75 | // (consistent with the RECSTORE_RDMA_* convention used by the raw-verbs path). | ||
| 76 | // RECSTORE_BRPC_USE_RDMA=1 -> enable RDMA transport | ||
| 77 | // RECSTORE_BRPC_RDMA_DEVICE=mlx5_0 -> select the HCA (maps to brpc | ||
| 78 | // -rdma_device) RECSTORE_BRPC_TIMEOUT_MS=60000 -> request timeout (maps to | ||
| 79 | // -brpc_timeout_ms) | ||
| 80 | 30 | bool ResolveBrpcUseRdmaFromEnv(bool fallback) { | |
| 81 | 30 | const char* value = std::getenv("RECSTORE_BRPC_USE_RDMA"); | |
| 82 | 30 | if (value == nullptr || *value == '\0') { | |
| 83 | 30 | return fallback; | |
| 84 | } | ||
| 85 | ✗ | return std::string(value) != "0"; | |
| 86 | } | ||
| 87 | |||
| 88 | ✗ | void ApplyBrpcRdmaDeviceFromEnv() { | |
| 89 | ✗ | const char* dev = std::getenv("RECSTORE_BRPC_RDMA_DEVICE"); | |
| 90 | ✗ | if (dev != nullptr && *dev != '\0') { | |
| 91 | ✗ | google::SetCommandLineOption("rdma_device", dev); | |
| 92 | } | ||
| 93 | ✗ | } | |
| 94 | |||
| 95 | 30 | int BrpcTimeoutMsFromEnv(int fallback) { | |
| 96 | 30 | const char* ms = std::getenv("RECSTORE_BRPC_TIMEOUT_MS"); | |
| 97 | 30 | if (ms == nullptr || *ms == '\0') { | |
| 98 | 30 | return fallback; | |
| 99 | } | ||
| 100 | try { | ||
| 101 | ✗ | return std::stoi(ms); | |
| 102 | ✗ | } catch (...) { | |
| 103 | ✗ | return fallback; | |
| 104 | ✗ | } | |
| 105 | } | ||
| 106 | |||
| 107 | } // namespace | ||
| 108 | |||
| 109 | // New constructor that takes JSON config | ||
| 110 | 24 | BRPCParameterClient::BRPCParameterClient(json config) { | |
| 111 | 24 | host_ = config.value("host", "localhost"); | |
| 112 | 24 | port_ = config.value("port", 15000); | |
| 113 | 24 | shard_ = config.value("shard", 0); | |
| 114 | 24 | timeout_ms_ = | |
| 115 | 24 | BrpcTimeoutMsFromEnv(config.value("timeout_ms", FLAGS_brpc_timeout_ms)); | |
| 116 | 24 | max_retry_ = config.value("max_retry", FLAGS_brpc_max_retry); | |
| 117 | |||
| 118 | 24 | Initialize(); | |
| 119 | |||
| 120 | // Initialize bRPC channel | ||
| 121 | 24 | channel_ = std::make_shared<brpc::Channel>(); | |
| 122 | 24 | brpc::ChannelOptions options; | |
| 123 | 24 | options.timeout_ms = timeout_ms_; | |
| 124 | 24 | options.max_retry = max_retry_; | |
| 125 | 48 | bool use_rdma_enabled = ResolveBrpcUseRdmaFromEnv( | |
| 126 | 24 | config.value("use_rdma", FLAGS_brpc_ps_use_rdma)); | |
| 127 | 24 | if (use_rdma_enabled) { | |
| 128 | ✗ | ApplyBrpcRdmaDeviceFromEnv(); | |
| 129 | } | ||
| 130 | #if BRPC_WITH_RDMA | ||
| 131 | options.use_rdma = use_rdma_enabled; | ||
| 132 | #endif | ||
| 133 | |||
| 134 | 24 | std::string server_addr = fmt::format("{}:{}", host_, port_); | |
| 135 | 24 | if (channel_->Init(server_addr.c_str(), &options) != 0) { | |
| 136 | ✗ | LOG(ERROR) << "Failed to initialize bRPC channel to " << server_addr; | |
| 137 | } else { | ||
| 138 | 48 | LOG(INFO) << "Initialized bRPC PS Client Shard " << shard_ << " at " | |
| 139 | 24 | << server_addr << " (use_rdma=" << use_rdma_enabled << ")"; | |
| 140 | } | ||
| 141 | 24 | } | |
| 142 | |||
| 143 | // Legacy constructor for backward compatibility | ||
| 144 | 6 | BRPCParameterClient::BRPCParameterClient( | |
| 145 | 6 | const std::string& host, int port, int shard) | |
| 146 | 6 | : host_(host), | |
| 147 | 6 | port_(port), | |
| 148 | 6 | shard_(shard), | |
| 149 | 6 | timeout_ms_(BrpcTimeoutMsFromEnv(FLAGS_brpc_timeout_ms)), | |
| 150 | 12 | max_retry_(FLAGS_brpc_max_retry) { | |
| 151 | 6 | Initialize(); | |
| 152 | |||
| 153 | 6 | channel_ = std::make_shared<brpc::Channel>(); | |
| 154 | 6 | brpc::ChannelOptions options; | |
| 155 | 6 | options.timeout_ms = timeout_ms_; | |
| 156 | 6 | options.max_retry = max_retry_; | |
| 157 | 6 | bool use_rdma_enabled = ResolveBrpcUseRdmaFromEnv(FLAGS_brpc_ps_use_rdma); | |
| 158 | 6 | if (use_rdma_enabled) { | |
| 159 | ✗ | ApplyBrpcRdmaDeviceFromEnv(); | |
| 160 | } | ||
| 161 | #if BRPC_WITH_RDMA | ||
| 162 | options.use_rdma = use_rdma_enabled; | ||
| 163 | #endif | ||
| 164 | |||
| 165 | ✗ | std::string server_addr = fmt::format("{}:{}", host, port); | |
| 166 | 6 | if (channel_->Init(server_addr.c_str(), &options) != 0) { | |
| 167 | ✗ | LOG(ERROR) << "Failed to initialize bRPC channel to " << server_addr; | |
| 168 | } else { | ||
| 169 | 12 | LOG(INFO) << "Initialized bRPC PS Client Shard " << shard_ << " at " | |
| 170 | 6 | << server_addr << " (use_rdma=" << use_rdma_enabled << ")"; | |
| 171 | } | ||
| 172 | 6 | } | |
| 173 | |||
| 174 | 30 | bool BRPCParameterClient::Initialize() { return true; } | |
| 175 | |||
| 176 | 40 | int BRPCParameterClient::GetParameter(const base::ConstArray<uint64_t>& keys, | |
| 177 | base::RecTensor& values) { | ||
| 178 | #ifdef ENABLE_PERF_REPORT | ||
| 179 | auto start_time = std::chrono::high_resolution_clock::now(); | ||
| 180 | #endif | ||
| 181 | |||
| 182 | 40 | if (!recstore::IsFloatEmbeddingValues( | |
| 183 | 40 | values, static_cast<int64_t>(keys.Size()))) { | |
| 184 | ✗ | return -1; | |
| 185 | } | ||
| 186 | 40 | if (keys.Size() == 0) { | |
| 187 | ✗ | return 0; | |
| 188 | } | ||
| 189 | 40 | const int64_t D = values.shape(1); | |
| 190 | 40 | float* dst = values.data_as<float>(); | |
| 191 | 40 | std::memset(dst, 0, keys.Size() * static_cast<size_t>(D) * sizeof(float)); | |
| 192 | |||
| 193 | 40 | if (FLAGS_parameter_client_random_init_brpc) { | |
| 194 | ✗ | std::fill(dst, dst + keys.Size() * static_cast<size_t>(D), 0.1f); | |
| 195 | ✗ | return 0; | |
| 196 | } | ||
| 197 | |||
| 198 | int request_num = | ||
| 199 | 40 | (keys.Size() + MAX_PARAMETER_BATCH_BRPC - 1) / MAX_PARAMETER_BATCH_BRPC; | |
| 200 | 40 | google::protobuf::Arena arena; | |
| 201 | 40 | std::vector<GetParameterRequest*> requests(request_num); | |
| 202 | 40 | std::vector<GetParameterResponse*> responses(request_num); | |
| 203 | 80 | for (int i = 0; i < request_num; ++i) { | |
| 204 | 40 | requests[i] = google::protobuf::Arena::Create<GetParameterRequest>(&arena); | |
| 205 | 40 | responses[i] = | |
| 206 | 40 | google::protobuf::Arena::Create<GetParameterResponse>(&arena); | |
| 207 | } | ||
| 208 | 40 | std::vector<brpc::Controller> controllers(request_num); | |
| 209 | 40 | std::vector<int> key_sizes; | |
| 210 | |||
| 211 | 40 | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 212 | |||
| 213 | #ifdef ENABLE_PERF_REPORT | ||
| 214 | auto wait_start_time = std::chrono::high_resolution_clock::now(); | ||
| 215 | #endif | ||
| 216 | |||
| 217 | // Send async RPC requests | ||
| 218 | 80 | for (int start = 0, index = 0; start < keys.Size(); | |
| 219 | 40 | start += MAX_PARAMETER_BATCH_BRPC, ++index) { | |
| 220 | int key_size = | ||
| 221 | 40 | std::min((int)(keys.Size() - start), MAX_PARAMETER_BATCH_BRPC); | |
| 222 | 40 | key_sizes.push_back(key_size); | |
| 223 | |||
| 224 | 80 | controllers[index].request_attachment().append( | |
| 225 | 40 | reinterpret_cast<const char*>(&keys[start]), | |
| 226 | 40 | sizeof(uint64_t) * key_size); | |
| 227 | |||
| 228 | 40 | google::protobuf::Closure* done = brpc::NewCallback([]() { /* no-op */ }); | |
| 229 | 40 | stub.GetParameter( | |
| 230 | 40 | &controllers[index], requests[index], responses[index], done); | |
| 231 | } | ||
| 232 | |||
| 233 | // Wait for all RPCs to complete | ||
| 234 | 80 | for (int i = 0; i < request_num; ++i) { | |
| 235 | 40 | brpc::Join(controllers[i].call_id()); | |
| 236 | 40 | if (controllers[i].Failed()) { | |
| 237 | ✗ | LOG(ERROR) << "bRPC GetParameter failed: " << controllers[i].ErrorText(); | |
| 238 | ✗ | return -1; | |
| 239 | } | ||
| 240 | } | ||
| 241 | |||
| 242 | #ifdef ENABLE_PERF_REPORT | ||
| 243 | auto wait_end_time = std::chrono::high_resolution_clock::now(); | ||
| 244 | auto wait_duration = | ||
| 245 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 246 | wait_end_time - wait_start_time) | ||
| 247 | .count(); | ||
| 248 | double wait_start_us = | ||
| 249 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 250 | wait_start_time.time_since_epoch()) | ||
| 251 | .count(); | ||
| 252 | std::string wait_label = | ||
| 253 | "brpc_client::RPC_Call_And_Wait_Shard" + std::to_string(shard_); | ||
| 254 | FlameGraphData wait_fg = { | ||
| 255 | wait_label, | ||
| 256 | wait_start_us, | ||
| 257 | 2, // level | ||
| 258 | static_cast<double>(wait_duration), | ||
| 259 | static_cast<double>(wait_duration)}; | ||
| 260 | std::string unique_id = | ||
| 261 | "embread_debug|" + std::to_string(static_cast<uint64_t>(wait_start_us)); | ||
| 262 | report_flame_graph("emb_read_flame_map", unique_id.c_str(), wait_fg); | ||
| 263 | |||
| 264 | double start_us_for_rpc = | ||
| 265 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 266 | start_time.time_since_epoch()) | ||
| 267 | .count(); | ||
| 268 | std::string report_id_for_rpc = | ||
| 269 | "brpc_client::GetParameter_Vec|" + | ||
| 270 | std::to_string(static_cast<uint64_t>(start_us_for_rpc)); | ||
| 271 | |||
| 272 | report("embread_stages", | ||
| 273 | report_id_for_rpc.c_str(), | ||
| 274 | "rpc_duration_us", | ||
| 275 | static_cast<double>(wait_duration)); | ||
| 276 | |||
| 277 | auto deserialize_start_time = std::chrono::high_resolution_clock::now(); | ||
| 278 | #endif | ||
| 279 | |||
| 280 | // Parse responses | ||
| 281 | 40 | std::string payload_storage; | |
| 282 | 40 | size_t row_offset = 0; | |
| 283 | 80 | for (int i = 0; i < responses.size(); ++i) { | |
| 284 | 40 | auto* response = responses[i]; | |
| 285 | 40 | int key_size = key_sizes[i]; | |
| 286 | 40 | int payload_size = 0; | |
| 287 | 40 | auto parameters = ExtractGetResponseReader( | |
| 288 | 40 | controllers[i], *response, &payload_storage, &payload_size); | |
| 289 | |||
| 290 | 40 | if (payload_size == 0 || parameters == nullptr) { | |
| 291 | ✗ | row_offset += static_cast<size_t>(key_size); | |
| 292 | ✗ | continue; | |
| 293 | } | ||
| 294 | 40 | if (!parameters->Valid(payload_size)) { | |
| 295 | ✗ | LOG(ERROR) << "GetParameter invalid payload: " << payload_size; | |
| 296 | ✗ | return -1; | |
| 297 | } | ||
| 298 | |||
| 299 | 40 | if (parameters->size == 0) { | |
| 300 | ✗ | row_offset += static_cast<size_t>(key_size); | |
| 301 | ✗ | continue; | |
| 302 | } | ||
| 303 | 40 | if (unlikely(parameters->size != key_size)) { | |
| 304 | ✗ | LOG(ERROR) << "GetParameter error: " << parameters->size << " vs " | |
| 305 | ✗ | << key_size; | |
| 306 | ✗ | return -1; | |
| 307 | } | ||
| 308 | |||
| 309 | 40 | CopyCompressItemsToFlat(parameters, dst, D, row_offset); | |
| 310 | 40 | row_offset += static_cast<size_t>(key_size); | |
| 311 | } | ||
| 312 | |||
| 313 | #ifdef ENABLE_PERF_REPORT | ||
| 314 | auto deserialize_end_time = std::chrono::high_resolution_clock::now(); | ||
| 315 | auto deserialize_duration = | ||
| 316 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 317 | deserialize_end_time - deserialize_start_time) | ||
| 318 | .count(); | ||
| 319 | double deserialize_start_us = | ||
| 320 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 321 | deserialize_start_time.time_since_epoch()) | ||
| 322 | .count(); | ||
| 323 | std::string des_label = | ||
| 324 | "brpc_client::Deserialize_Shard" + std::to_string(shard_); | ||
| 325 | FlameGraphData des_fg = { | ||
| 326 | des_label, | ||
| 327 | deserialize_start_us, | ||
| 328 | 2, // level | ||
| 329 | static_cast<double>(deserialize_duration), | ||
| 330 | static_cast<double>(deserialize_duration)}; | ||
| 331 | std::string des_unique_id = | ||
| 332 | "embread_debug|" + | ||
| 333 | std::to_string(static_cast<uint64_t>(deserialize_start_us)); | ||
| 334 | report_flame_graph("emb_read_flame_map", des_unique_id.c_str(), des_fg); | ||
| 335 | |||
| 336 | double start_us_for_des = | ||
| 337 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 338 | start_time.time_since_epoch()) | ||
| 339 | .count(); | ||
| 340 | std::string report_id_for_des = | ||
| 341 | "brpc_client::GetParameter_Vec|" + | ||
| 342 | std::to_string(static_cast<uint64_t>(start_us_for_des)); | ||
| 343 | report("embread_stages", | ||
| 344 | report_id_for_des.c_str(), | ||
| 345 | "deserialize_duration_us", | ||
| 346 | static_cast<double>(deserialize_duration)); | ||
| 347 | #endif | ||
| 348 | |||
| 349 | #ifdef ENABLE_PERF_REPORT | ||
| 350 | auto end_time = std::chrono::high_resolution_clock::now(); | ||
| 351 | auto duration = | ||
| 352 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 353 | end_time - start_time) | ||
| 354 | .count(); | ||
| 355 | report("ps_client_latency", | ||
| 356 | "GetParameter", | ||
| 357 | "latency_us", | ||
| 358 | static_cast<double>(duration)); | ||
| 359 | |||
| 360 | double start_us = | ||
| 361 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 362 | start_time.time_since_epoch()) | ||
| 363 | .count(); | ||
| 364 | FlameGraphData fg_data = { | ||
| 365 | "brpc_client::GetParameter_Vec", | ||
| 366 | start_us, | ||
| 367 | 1, // level | ||
| 368 | static_cast<double>(duration), | ||
| 369 | static_cast<double>(duration)}; | ||
| 370 | |||
| 371 | std::string report_id = "brpc_client::GetParameter_Vec|" + | ||
| 372 | std::to_string(static_cast<uint64_t>(start_us)); | ||
| 373 | |||
| 374 | report("embread_stages", | ||
| 375 | report_id.c_str(), | ||
| 376 | "duration_us", | ||
| 377 | static_cast<double>(duration)); | ||
| 378 | |||
| 379 | report("embread_stages", | ||
| 380 | report_id.c_str(), | ||
| 381 | "request_size", | ||
| 382 | static_cast<double>(keys.Size())); | ||
| 383 | |||
| 384 | std::string final_unique_id = | ||
| 385 | "embread_debug|" + std::to_string(static_cast<uint64_t>(start_us)); | ||
| 386 | report_flame_graph("emb_read_flame_map", final_unique_id.c_str(), fg_data); | ||
| 387 | #endif | ||
| 388 | |||
| 389 | 40 | return 0; | |
| 390 | 40 | } | |
| 391 | |||
| 392 | 10 | static void OnPrefetchDone(BrpcPrefetchBatch* batch) { | |
| 393 | 10 | batch->completed_count_++; | |
| 394 | 10 | } | |
| 395 | |||
| 396 | uint64_t | ||
| 397 | 10 | BRPCParameterClient::PrefetchParameter(const base::ConstArray<uint64_t>& keys) { | |
| 398 | 10 | uint64_t prefetch_id = next_prefetch_id_++; | |
| 399 | int request_num = | ||
| 400 | 10 | (keys.Size() + MAX_PARAMETER_BATCH_BRPC - 1) / MAX_PARAMETER_BATCH_BRPC; | |
| 401 | |||
| 402 | // Construct in map so batch pointers stay valid | ||
| 403 | 10 | auto it = prefetch_batches_.emplace(prefetch_id, request_num).first; | |
| 404 | 10 | struct BrpcPrefetchBatch* pb = &it->second; | |
| 405 | |||
| 406 | 10 | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 407 | |||
| 408 | 20 | for (int start = 0, index = 0; start < keys.Size(); | |
| 409 | 10 | start += MAX_PARAMETER_BATCH_BRPC, ++index) { | |
| 410 | int key_size = | ||
| 411 | 10 | std::min((int)(keys.Size() - start), MAX_PARAMETER_BATCH_BRPC); | |
| 412 | 10 | pb->key_sizes_[index] = key_size; | |
| 413 | |||
| 414 | 10 | GetParameterRequest request; | |
| 415 | |||
| 416 | 10 | pb->controllers_[index] = std::make_unique<brpc::Controller>(); | |
| 417 | 20 | pb->controllers_[index]->request_attachment().append( | |
| 418 | 10 | reinterpret_cast<const char*>(&keys[start]), | |
| 419 | 10 | sizeof(uint64_t) * key_size); | |
| 420 | |||
| 421 | 10 | google::protobuf::Closure* done = brpc::NewCallback(OnPrefetchDone, pb); | |
| 422 | 10 | stub.GetParameter( | |
| 423 | 10 | pb->controllers_[index].get(), &request, &pb->responses_[index], done); | |
| 424 | 10 | } | |
| 425 | |||
| 426 | 10 | return prefetch_id; | |
| 427 | 10 | } | |
| 428 | |||
| 429 | ✗ | bool BRPCParameterClient::IsPrefetchDone(uint64_t prefetch_id) { | |
| 430 | ✗ | auto it = prefetch_batches_.find(prefetch_id); | |
| 431 | ✗ | if (it == prefetch_batches_.end()) { | |
| 432 | ✗ | LOG(ERROR) << "Invalid prefetch_id: " << prefetch_id; | |
| 433 | ✗ | return false; | |
| 434 | } | ||
| 435 | |||
| 436 | ✗ | auto& pb = it->second; | |
| 437 | |||
| 438 | ✗ | return pb.completed_count_ == pb.batch_size_; | |
| 439 | } | ||
| 440 | |||
| 441 | 10 | void BRPCParameterClient::WaitForPrefetch(uint64_t prefetch_id) { | |
| 442 | 10 | auto it = prefetch_batches_.find(prefetch_id); | |
| 443 | 10 | if (it == prefetch_batches_.end()) { | |
| 444 | ✗ | LOG(ERROR) << "Invalid prefetch_id: " << prefetch_id; | |
| 445 | ✗ | return; | |
| 446 | } | ||
| 447 | 10 | auto& pb = it->second; | |
| 448 | 20 | for (int i = 0; i < pb.batch_size_; ++i) { | |
| 449 | 10 | if (pb.controllers_[i]) { | |
| 450 | 10 | brpc::Join(pb.controllers_[i]->call_id()); | |
| 451 | } | ||
| 452 | } | ||
| 453 | 10 | pb.completed_count_ = pb.batch_size_; | |
| 454 | } | ||
| 455 | |||
| 456 | 10 | bool BRPCParameterClient::GetPrefetchResult(uint64_t prefetch_id, | |
| 457 | base::RecTensor& values) { | ||
| 458 | 10 | auto it = prefetch_batches_.find(prefetch_id); | |
| 459 | 10 | if (it == prefetch_batches_.end()) { | |
| 460 | ✗ | LOG(ERROR) << "Invalid prefetch_id: " << prefetch_id; | |
| 461 | ✗ | return false; | |
| 462 | } | ||
| 463 | |||
| 464 | 10 | auto& pb = it->second; | |
| 465 | 10 | int request_num = pb.batch_size_; | |
| 466 | |||
| 467 | 10 | int keys_size = 0; | |
| 468 | 20 | for (const auto& size : pb.key_sizes_) { | |
| 469 | 10 | keys_size += size; | |
| 470 | } | ||
| 471 | 10 | const bool discard = values.data() == nullptr && values.dim() == 0; | |
| 472 | 10 | if (!discard && !recstore::EnsureEmbeddingOutput( | |
| 473 | values, static_cast<int64_t>(keys_size))) { | ||
| 474 | ✗ | return false; | |
| 475 | } | ||
| 476 | 10 | const int64_t D = discard ? 0 : values.shape(1); | |
| 477 | 10 | float* dst = discard ? nullptr : values.data_as<float>(); | |
| 478 | 10 | if (!discard && dst != nullptr && keys_size > 0) { | |
| 479 | 10 | std::memset(dst, | |
| 480 | 0, | ||
| 481 | 10 | static_cast<size_t>(keys_size) * static_cast<size_t>(D) * | |
| 482 | sizeof(float)); | ||
| 483 | } | ||
| 484 | 10 | size_t row_offset = 0; | |
| 485 | |||
| 486 | 20 | for (int i = 0; i < request_num; ++i) { | |
| 487 | 10 | if (pb.controllers_[i]->Failed()) { | |
| 488 | ✗ | LOG(ERROR) << "Prefetch request failed: " | |
| 489 | ✗ | << pb.controllers_[i]->ErrorText(); | |
| 490 | ✗ | return false; | |
| 491 | } | ||
| 492 | |||
| 493 | 10 | auto& response = pb.responses_[i]; | |
| 494 | 10 | int key_size = pb.key_sizes_[i]; | |
| 495 | 10 | std::string payload_storage; | |
| 496 | 10 | int payload_size = 0; | |
| 497 | 10 | auto parameters = ExtractGetResponseReader( | |
| 498 | 10 | *pb.controllers_[i], response, &payload_storage, &payload_size); | |
| 499 | |||
| 500 | 10 | if (payload_size == 0 || parameters == nullptr) { | |
| 501 | ✗ | row_offset += static_cast<size_t>(key_size); | |
| 502 | ✗ | continue; | |
| 503 | } | ||
| 504 | 10 | if (!parameters->Valid(payload_size)) { | |
| 505 | ✗ | LOG(ERROR) << "Prefetch invalid payload: " << payload_size; | |
| 506 | ✗ | return false; | |
| 507 | } | ||
| 508 | |||
| 509 | 10 | if (parameters->size == 0) { | |
| 510 | ✗ | row_offset += static_cast<size_t>(key_size); | |
| 511 | ✗ | continue; | |
| 512 | } | ||
| 513 | 10 | if (unlikely(parameters->size != key_size)) { | |
| 514 | ✗ | LOG(ERROR) << "GetParameter error: " << parameters->size << " vs " | |
| 515 | ✗ | << key_size; | |
| 516 | ✗ | return false; | |
| 517 | } | ||
| 518 | |||
| 519 | 10 | if (!discard) { | |
| 520 | 10 | CopyCompressItemsToFlat(parameters, dst, D, row_offset); | |
| 521 | 10 | row_offset += static_cast<size_t>(key_size); | |
| 522 | } | ||
| 523 | 10 | } | |
| 524 | |||
| 525 | 10 | prefetch_batches_.erase(it); | |
| 526 | |||
| 527 | 10 | return true; | |
| 528 | } | ||
| 529 | |||
| 530 | 34 | bool BRPCParameterClient::ClearPS() { | |
| 531 | 34 | CommandRequest request; | |
| 532 | 34 | CommandResponse response; | |
| 533 | 34 | request.set_command(PSCommand::CLEAR_PS); | |
| 534 | |||
| 535 | 34 | brpc::Controller cntl; | |
| 536 | 34 | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 537 | 34 | stub.Command(&cntl, &request, &response, nullptr); | |
| 538 | |||
| 539 | 34 | if (cntl.Failed()) { | |
| 540 | ✗ | LOG(ERROR) << "bRPC Command failed: " << cntl.ErrorText(); | |
| 541 | ✗ | return false; | |
| 542 | } | ||
| 543 | 34 | return true; | |
| 544 | 34 | } | |
| 545 | |||
| 546 | 6 | bool BRPCParameterClient::LoadFakeData(int64_t data) { | |
| 547 | 6 | CommandRequest request; | |
| 548 | 6 | CommandResponse response; | |
| 549 | 6 | request.set_command(PSCommand::LOAD_FAKE_DATA); | |
| 550 | 6 | request.add_arg1(&data, sizeof(int64_t)); | |
| 551 | |||
| 552 | 6 | brpc::Controller cntl; | |
| 553 | 6 | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 554 | 6 | stub.Command(&cntl, &request, &response, nullptr); | |
| 555 | |||
| 556 | 6 | if (cntl.Failed()) { | |
| 557 | ✗ | LOG(ERROR) << "bRPC LoadFakeData failed: " << cntl.ErrorText(); | |
| 558 | ✗ | return false; | |
| 559 | } | ||
| 560 | 6 | if (response.reply().size() != static_cast<size_t>(data)) { | |
| 561 | ✗ | LOG(ERROR) << "bRPC LoadFakeData reply size mismatch: expected " << data | |
| 562 | ✗ | << ", got " << response.reply().size(); | |
| 563 | ✗ | return false; | |
| 564 | } | ||
| 565 | 6 | return true; | |
| 566 | 6 | } | |
| 567 | |||
| 568 | 6 | bool BRPCParameterClient::DumpFakeData(int64_t n) { | |
| 569 | 6 | CommandRequest request; | |
| 570 | 6 | CommandResponse response; | |
| 571 | 6 | request.set_command(PSCommand::DUMP_FAKE_DATA); | |
| 572 | 6 | request.add_arg1(&n, sizeof(int64_t)); | |
| 573 | |||
| 574 | 6 | brpc::Controller cntl; | |
| 575 | 6 | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 576 | 6 | stub.Command(&cntl, &request, &response, nullptr); | |
| 577 | |||
| 578 | 6 | if (cntl.Failed()) { | |
| 579 | ✗ | LOG(ERROR) << "bRPC DumpFakeData failed: " << cntl.ErrorText(); | |
| 580 | ✗ | return false; | |
| 581 | } | ||
| 582 | 6 | if (response.reply() != "ok") { | |
| 583 | ✗ | LOG(ERROR) << "bRPC DumpFakeData unexpected reply: " << response.reply(); | |
| 584 | ✗ | return false; | |
| 585 | } | ||
| 586 | 6 | return true; | |
| 587 | 6 | } | |
| 588 | |||
| 589 | ✗ | bool BRPCParameterClient::LoadCkpt( | |
| 590 | const std::vector<std::string>& model_config_path, | ||
| 591 | const std::vector<std::string>& emb_file_path) { | ||
| 592 | ✗ | CommandRequest request; | |
| 593 | ✗ | CommandResponse response; | |
| 594 | ✗ | request.set_command(PSCommand::RELOAD_PS); | |
| 595 | |||
| 596 | ✗ | for (auto& each : model_config_path) { | |
| 597 | ✗ | request.add_arg1(each); | |
| 598 | } | ||
| 599 | ✗ | for (auto& each : emb_file_path) { | |
| 600 | ✗ | request.add_arg2(each); | |
| 601 | } | ||
| 602 | |||
| 603 | ✗ | brpc::Controller cntl; | |
| 604 | ✗ | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 605 | ✗ | stub.Command(&cntl, &request, &response, nullptr); | |
| 606 | |||
| 607 | ✗ | if (cntl.Failed()) { | |
| 608 | ✗ | LOG(ERROR) << "bRPC LoadCkpt failed: " << cntl.ErrorText(); | |
| 609 | ✗ | return false; | |
| 610 | } | ||
| 611 | ✗ | return true; | |
| 612 | ✗ | } | |
| 613 | |||
| 614 | 28 | bool BRPCParameterClient::PutParameter(const std::vector<uint64_t>& keys, | |
| 615 | const base::RecTensor& values) { | ||
| 616 | #ifdef ENABLE_PERF_REPORT | ||
| 617 | auto start_time = std::chrono::high_resolution_clock::now(); | ||
| 618 | #endif | ||
| 619 | 28 | if (!recstore::IsFloatEmbeddingValues( | |
| 620 | 28 | values, static_cast<int64_t>(keys.size()))) { | |
| 621 | ✗ | return false; | |
| 622 | } | ||
| 623 | 28 | const int64_t D = keys.empty() ? 0 : values.shape(1); | |
| 624 | 28 | const float* src = keys.empty() ? nullptr : values.data_as<float>(); | |
| 625 | |||
| 626 | 28 | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 627 | |||
| 628 | 56 | for (int start = 0, index = 0; start < keys.size(); | |
| 629 | 28 | start += MAX_PARAMETER_BATCH_BRPC, ++index) { | |
| 630 | int key_size = | ||
| 631 | 28 | std::min((int)(keys.size() - start), MAX_PARAMETER_BATCH_BRPC); | |
| 632 | |||
| 633 | 28 | PutParameterRequest request; | |
| 634 | 28 | PutParameterResponse response; | |
| 635 | 28 | ParameterCompressor compressor; | |
| 636 | 56 | CompressEmbeddingRows( | |
| 637 | 28 | &compressor, keys.data() + start, src + start * D, key_size, D); | |
| 638 | |||
| 639 | 28 | brpc::Controller cntl; | |
| 640 | 28 | compressor.AppendToIOBuf(&cntl.request_attachment()); | |
| 641 | 28 | stub.PutParameter(&cntl, &request, &response, nullptr); | |
| 642 | |||
| 643 | 28 | if (cntl.Failed()) { | |
| 644 | ✗ | LOG(ERROR) << "bRPC PutParameter failed: " << cntl.ErrorText(); | |
| 645 | ✗ | return false; | |
| 646 | } | ||
| 647 | 28 | } | |
| 648 | |||
| 649 | #ifdef ENABLE_PERF_REPORT | ||
| 650 | auto end_time = std::chrono::high_resolution_clock::now(); | ||
| 651 | auto duration = | ||
| 652 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 653 | end_time - start_time) | ||
| 654 | .count(); | ||
| 655 | report("ps_client_latency", | ||
| 656 | "PutParameter", | ||
| 657 | "latency_us", | ||
| 658 | static_cast<double>(duration)); | ||
| 659 | #endif | ||
| 660 | |||
| 661 | 28 | return true; | |
| 662 | 28 | } | |
| 663 | |||
| 664 | 14 | int BRPCParameterClient::PutParameter(const base::ConstArray<uint64_t>& keys, | |
| 665 | const base::RecTensor& values) { | ||
| 666 | 14 | std::vector<uint64_t> key_vec(keys.Data(), keys.Data() + keys.Size()); | |
| 667 | 14 | bool success = PutParameter(key_vec, values); | |
| 668 | 28 | return success ? 0 : -1; | |
| 669 | 14 | } | |
| 670 | |||
| 671 | ✗ | void BRPCParameterClient::Command(recstore::PSCommand command) { | |
| 672 | ✗ | switch (command) { | |
| 673 | ✗ | case recstore::PSCommand::CLEAR_PS: | |
| 674 | ✗ | ClearPS(); | |
| 675 | ✗ | break; | |
| 676 | ✗ | case recstore::PSCommand::RELOAD_PS: | |
| 677 | ✗ | LOG(WARNING) << "RELOAD_PS command requires additional parameters"; | |
| 678 | ✗ | break; | |
| 679 | ✗ | case recstore::PSCommand::LOAD_FAKE_DATA: { | |
| 680 | ✗ | int64_t fake_data = 1000; | |
| 681 | ✗ | LoadFakeData(fake_data); | |
| 682 | ✗ | } break; | |
| 683 | ✗ | case recstore::PSCommand::DUMP_FAKE_DATA: { | |
| 684 | ✗ | DumpFakeData(4096); | |
| 685 | ✗ | } break; | |
| 686 | ✗ | default: | |
| 687 | ✗ | LOG(ERROR) << "Unknown PS command: " << static_cast<int>(command); | |
| 688 | ✗ | break; | |
| 689 | } | ||
| 690 | ✗ | } | |
| 691 | |||
| 692 | ✗ | int BRPCParameterClient::UpdateParameter(const std::string& table_name, | |
| 693 | const base::ConstArray<uint64_t>& keys, | ||
| 694 | const base::RecTensor& grads) { | ||
| 695 | #ifdef ENABLE_PERF_REPORT | ||
| 696 | auto start_time = std::chrono::high_resolution_clock::now(); | ||
| 697 | const uint64_t trace_id = recstore::g_trace_id; | ||
| 698 | #endif | ||
| 699 | ✗ | if (!recstore::IsFloatEmbeddingValues( | |
| 700 | ✗ | grads, static_cast<int64_t>(keys.Size()))) { | |
| 701 | ✗ | LOG(ERROR) << "UpdateParameter keys/grads size mismatch: " << keys.Size(); | |
| 702 | ✗ | return -1; | |
| 703 | } | ||
| 704 | |||
| 705 | ✗ | ParameterCompressor compressor; | |
| 706 | ✗ | if (keys.Size() > 0) { | |
| 707 | ✗ | CompressEmbeddingRows( | |
| 708 | &compressor, | ||
| 709 | keys.Data(), | ||
| 710 | ✗ | grads.data_as<float>(), | |
| 711 | ✗ | keys.Size(), | |
| 712 | grads.shape(1)); | ||
| 713 | } | ||
| 714 | #ifdef ENABLE_PERF_REPORT | ||
| 715 | auto serialize_done_time = std::chrono::high_resolution_clock::now(); | ||
| 716 | #endif | ||
| 717 | ✗ | if (keys.Size() == 0) { | |
| 718 | ✗ | LOG(WARNING) << "UpdateParameter no gradients to send"; | |
| 719 | ✗ | return 0; | |
| 720 | } | ||
| 721 | |||
| 722 | ✗ | UpdateParameterRequest request; | |
| 723 | ✗ | UpdateParameterResponse response; | |
| 724 | request.set_table_name(table_name); | ||
| 725 | |||
| 726 | ✗ | brpc::Controller cntl; | |
| 727 | #ifdef ENABLE_PERF_REPORT | ||
| 728 | if (trace_id != 0) { | ||
| 729 | cntl.http_request().SetHeader( | ||
| 730 | "x-recstore-trace-id", std::to_string(trace_id)); | ||
| 731 | } | ||
| 732 | auto rpc_start_time = std::chrono::high_resolution_clock::now(); | ||
| 733 | #endif | ||
| 734 | ✗ | compressor.AppendToIOBuf(&cntl.request_attachment()); | |
| 735 | ✗ | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 736 | ✗ | stub.UpdateParameter(&cntl, &request, &response, nullptr); | |
| 737 | ✗ | if (cntl.Failed()) { | |
| 738 | ✗ | LOG(ERROR) << "UpdateParameter RPC failed: " << cntl.ErrorText(); | |
| 739 | ✗ | return -1; | |
| 740 | } | ||
| 741 | |||
| 742 | #ifdef ENABLE_PERF_REPORT | ||
| 743 | auto end_time = std::chrono::high_resolution_clock::now(); | ||
| 744 | auto duration = | ||
| 745 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 746 | end_time - start_time) | ||
| 747 | .count(); | ||
| 748 | auto serialize_duration = | ||
| 749 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 750 | serialize_done_time - start_time) | ||
| 751 | .count(); | ||
| 752 | auto rpc_duration = | ||
| 753 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 754 | end_time - rpc_start_time) | ||
| 755 | .count(); | ||
| 756 | report("ps_client_latency", | ||
| 757 | "UpdateParameter", | ||
| 758 | "latency_us", | ||
| 759 | static_cast<double>(duration)); | ||
| 760 | |||
| 761 | const uint64_t effective_trace_id = | ||
| 762 | trace_id == 0 | ||
| 763 | ? static_cast<uint64_t>( | ||
| 764 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 765 | start_time.time_since_epoch()) | ||
| 766 | .count()) | ||
| 767 | : trace_id; | ||
| 768 | std::string stage_id = | ||
| 769 | "brpc_client::EmbUpdate|" + std::to_string(effective_trace_id); | ||
| 770 | report("embupdate_stages", | ||
| 771 | stage_id.c_str(), | ||
| 772 | "client_serialize_us", | ||
| 773 | static_cast<double>(serialize_duration)); | ||
| 774 | report("embupdate_stages", | ||
| 775 | stage_id.c_str(), | ||
| 776 | "client_rpc_us", | ||
| 777 | static_cast<double>(rpc_duration)); | ||
| 778 | report("embupdate_stages", | ||
| 779 | stage_id.c_str(), | ||
| 780 | "client_total_us", | ||
| 781 | static_cast<double>(duration)); | ||
| 782 | report("embupdate_stages", | ||
| 783 | stage_id.c_str(), | ||
| 784 | "client_request_size", | ||
| 785 | static_cast<double>(keys.Size())); | ||
| 786 | #endif | ||
| 787 | |||
| 788 | ✗ | return response.success() ? 0 : -1; | |
| 789 | ✗ | } | |
| 790 | |||
| 791 | ✗ | int BRPCParameterClient::InitEmbeddingTable( | |
| 792 | const std::string& table_name, | ||
| 793 | const recstore::EmbeddingTableConfig& config) { | ||
| 794 | #ifdef ENABLE_PERF_REPORT | ||
| 795 | auto start_time = std::chrono::high_resolution_clock::now(); | ||
| 796 | #endif | ||
| 797 | |||
| 798 | ✗ | InitEmbeddingTableRequest request; | |
| 799 | ✗ | InitEmbeddingTableResponse response; | |
| 800 | request.set_table_name(table_name); | ||
| 801 | ✗ | request.set_config_payload(config.Serialize()); | |
| 802 | |||
| 803 | ✗ | brpc::Controller cntl; | |
| 804 | ✗ | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 805 | ✗ | stub.InitEmbeddingTable(&cntl, &request, &response, nullptr); | |
| 806 | ✗ | if (cntl.Failed()) { | |
| 807 | ✗ | LOG(ERROR) << "InitEmbeddingTable RPC failed: " << cntl.ErrorText(); | |
| 808 | ✗ | return -1; | |
| 809 | } | ||
| 810 | |||
| 811 | #ifdef ENABLE_PERF_REPORT | ||
| 812 | auto end_time = std::chrono::high_resolution_clock::now(); | ||
| 813 | auto duration = | ||
| 814 | std::chrono::duration_cast<std::chrono::microseconds>( | ||
| 815 | end_time - start_time) | ||
| 816 | .count(); | ||
| 817 | report("ps_client_latency", | ||
| 818 | "InitEmbeddingTable", | ||
| 819 | "latency_us", | ||
| 820 | static_cast<double>(duration)); | ||
| 821 | #endif | ||
| 822 | |||
| 823 | ✗ | return response.success() ? response.tag() : -1; | |
| 824 | ✗ | } | |
| 825 | |||
| 826 | ✗ | int64_t BRPCParameterClient::ExchangeSparseGradientCommit( | |
| 827 | const std::string& table, const std::string& payload) { | ||
| 828 | ✗ | recstoreps_brpc::SparseGradientCommitRequest request; | |
| 829 | ✗ | recstoreps_brpc::SparseGradientCommitResponse response; | |
| 830 | request.set_table_name(table); | ||
| 831 | request.set_payload(payload); | ||
| 832 | ✗ | brpc::Controller controller; | |
| 833 | ✗ | recstoreps_brpc::ParameterService_Stub stub(channel_.get()); | |
| 834 | ✗ | stub.SparseGradientCommit(&controller, &request, &response, nullptr); | |
| 835 | ✗ | if (controller.Failed() || !response.error().empty()) { | |
| 836 | ✗ | throw std::runtime_error( | |
| 837 | ✗ | "SparseGradientCommit: " + | |
| 838 | ✗ | (controller.Failed() ? controller.ErrorText() : response.error())); | |
| 839 | } | ||
| 840 | ✗ | return response.committed_step(); | |
| 841 | ✗ | } | |
| 842 |