00001 #include <chrono>
00002
00003 #define TRACE_NAME "DataReceiverManager"
00004 #include "artdaq/DAQdata/Globals.hh"
00005 #include "artdaq/DAQrate/DataReceiverManager.hh"
00006 #include "artdaq/TransferPlugins/MakeTransferPlugin.hh"
00007 #include "artdaq/TransferPlugins/detail/HostMap.hh"
00008 #include "cetlib_except/exception.h"
00009 #include <iomanip>
00010
00011 artdaq::DataReceiverManager::DataReceiverManager(const fhicl::ParameterSet& pset, std::shared_ptr<SharedMemoryEventManager> shm)
00012 : stop_requested_(false)
00013 , stop_requested_time_(0)
00014 , source_threads_()
00015 , source_plugins_()
00016 , source_metric_data_()
00017 , source_metric_send_time_()
00018 , enabled_sources_()
00019 , running_sources_()
00020 , recv_frag_count_()
00021 , recv_frag_size_()
00022 , recv_seq_count_()
00023 , receive_timeout_(pset.get<size_t>("receive_timeout_usec", 100000))
00024 , stop_timeout_ms_(pset.get<size_t>("stop_timeout_ms", 1500))
00025 , shm_manager_(shm)
00026 , non_reliable_mode_enabled_(pset.get<bool>("non_reliable_mode", false))
00027 , non_reliable_mode_retry_count_(pset.get<size_t>("non_reliable_mode_retry_count", -1))
00028 {
00029 TLOG(TLVL_DEBUG) << "Constructor";
00030 auto enabled_srcs = pset.get<std::vector<int>>("enabled_sources", std::vector<int>());
00031 auto enabled_srcs_empty = enabled_srcs.size() == 0;
00032
00033 if (non_reliable_mode_enabled_)
00034 {
00035 TLOG(TLVL_WARNING) << "DataReceiverManager is configured to drop data after " << non_reliable_mode_retry_count_
00036 << " failed attempts to put data into the SharedMemoryEventManager! If this is unexpected, please check your configuration!";
00037 }
00038
00039 if (enabled_srcs_empty)
00040 {
00041 TLOG(TLVL_INFO) << "enabled_sources not specified, assuming all sources enabled.";
00042 }
00043 else
00044 {
00045 for (auto& s : enabled_srcs)
00046 {
00047 enabled_sources_[s] = true;
00048 }
00049 }
00050
00051 hostMap_t host_map = MakeHostMap(pset);
00052 auto srcs = pset.get<fhicl::ParameterSet>("sources", fhicl::ParameterSet());
00053 for (auto& s : srcs.get_pset_names())
00054 {
00055 auto src_pset = srcs.get<fhicl::ParameterSet>(s);
00056 host_map = MakeHostMap(src_pset, 0, host_map);
00057 }
00058 auto host_map_pset = MakeHostMapPset(host_map);
00059 fhicl::ParameterSet srcs_mod;
00060 for (auto& s : srcs.get_pset_names())
00061 {
00062 auto src_pset = srcs.get<fhicl::ParameterSet>(s);
00063 src_pset.erase("host_map");
00064 src_pset.put<std::vector<fhicl::ParameterSet>>("host_map", host_map_pset);
00065 srcs_mod.put<fhicl::ParameterSet>(s, src_pset);
00066 }
00067
00068 for (auto& s : srcs_mod.get_pset_names())
00069 {
00070 try
00071 {
00072 auto transfer = std::unique_ptr<TransferInterface>(MakeTransferPlugin(srcs_mod, s,
00073 TransferInterface::Role::kReceive));
00074 auto source_rank = transfer->source_rank();
00075 if (enabled_srcs_empty) enabled_sources_[source_rank] = true;
00076 else if (!enabled_sources_.count(source_rank)) enabled_sources_[source_rank] = false;
00077 running_sources_[source_rank] = false;
00078 source_plugins_[source_rank] = std::move(transfer);
00079 source_metric_send_time_[source_rank] = std::chrono::steady_clock::now();
00080 source_metric_data_[source_rank] = source_metric_data();
00081 }
00082 catch (cet::exception ex)
00083 {
00084 TLOG(TLVL_WARNING) << "cet::exception caught while setting up source " << s << ": " << ex.what();
00085 }
00086 catch (std::exception ex)
00087 {
00088 TLOG(TLVL_WARNING) << "std::exception caught while setting up source " << s << ": " << ex.what();
00089 }
00090 catch (...)
00091 {
00092 TLOG(TLVL_WARNING) << "Non-cet exception caught while setting up source " << s << ".";
00093 }
00094 }
00095 if (srcs.get_pset_names().size() == 0)
00096 {
00097 TLOG(TLVL_ERROR) << "No sources configured!";
00098 }
00099 }
00100
00101 artdaq::DataReceiverManager::~DataReceiverManager()
00102 {
00103 TLOG(TLVL_TRACE) << "~DataReceiverManager: BEGIN";
00104 stop_threads();
00105 shm_manager_.reset();
00106 TLOG(TLVL_TRACE) << "Destructor END";
00107 }
00108
00109
00110 void artdaq::DataReceiverManager::start_threads()
00111 {
00112 stop_requested_ = false;
00113 if (shm_manager_) shm_manager_->setRequestMode(artdaq::detail::RequestMessageMode::Normal);
00114 for (auto& source : source_plugins_)
00115 {
00116 auto& rank = source.first;
00117 if (enabled_sources_.count(rank) && enabled_sources_[rank].load())
00118 {
00119 running_sources_[rank] = true;
00120 boost::thread::attributes attrs;
00121 attrs.set_stack_size(4096 * 500);
00122 source_threads_[rank] = boost::thread(attrs, boost::bind(&DataReceiverManager::runReceiver_, this, rank));
00123 }
00124 }
00125 }
00126
00127 void artdaq::DataReceiverManager::stop_threads()
00128 {
00129 TLOG(TLVL_TRACE) << "stop_threads: BEGIN: Setting stop_requested to true, frags=" << count() << ", bytes=" << byteCount();
00130
00131 stop_requested_time_ = TimeUtils::gettimeofday_us();
00132 stop_requested_ = true;
00133
00134 TLOG(TLVL_TRACE) << "stop_threads: Joining all threads";
00135 for (auto& s : source_threads_)
00136 {
00137 auto& thread = s.second;
00138 if (thread.joinable()) thread.join();
00139 }
00140 }
00141
00142 std::set<int> artdaq::DataReceiverManager::enabled_sources() const
00143 {
00144 std::set<int> output;
00145 for (auto& src : enabled_sources_)
00146 {
00147 if (src.second) output.insert(src.first);
00148 }
00149 return output;
00150 }
00151
00152 std::set<int> artdaq::DataReceiverManager::running_sources() const
00153 {
00154 std::set<int> output;
00155 for (auto& src : running_sources_)
00156 {
00157 if (src.second) output.insert(src.first);
00158 }
00159 return output;
00160 }
00161
00162 void artdaq::DataReceiverManager::runReceiver_(int source_rank)
00163 {
00164 std::chrono::steady_clock::time_point start_time, after_header, before_body;
00165 int ret;
00166 detail::RawFragmentHeader header;
00167 size_t endOfDataCount = -1;
00168 auto sleep_time = receive_timeout_ / 100 > 100000 ? 100000 : receive_timeout_ / 100;
00169 auto max_retries = non_reliable_mode_retry_count_ * ceil(receive_timeout_ / sleep_time);
00170
00171 while (!(stop_requested_ && TimeUtils::gettimeofday_us() - stop_requested_time_ > stop_timeout_ms_ * 1000) && enabled_sources_.count(source_rank))
00172 {
00173 TLOG(16) << "runReceiver_: Begin loop";
00174 std::this_thread::yield();
00175
00176
00177 if (endOfDataCount <= recv_frag_count_.slotCount(source_rank) && !source_plugins_[source_rank]->isRunning())
00178 {
00179 TLOG(TLVL_DEBUG) << "runReceiver_: End of Data conditions met, ending runReceiver loop";
00180 break;
00181 }
00182
00183 start_time = std::chrono::steady_clock::now();
00184
00185 TLOG(16) << "runReceiver_: Calling receiveFragmentHeader tmo=" << receive_timeout_;
00186 ret = source_plugins_[source_rank]->receiveFragmentHeader(header, receive_timeout_);
00187 TLOG(16) << "runReceiver_: Done with receiveFragmentHeader, ret=" << ret << " (should be " << source_rank << ")";
00188 if (ret != source_rank)
00189 {
00190 if (ret >= 0) {
00191 TLOG(TLVL_WARNING) << "Received Fragment from rank " << ret << ", but was expecting one from rank " << source_rank << "!";
00192 }
00193 else if(ret == TransferInterface::DATA_END)
00194 {
00195 TLOG(TLVL_ERROR) << "Transfer Plugin returned DATA_END, ending receive loop!";
00196 break;
00197 }
00198 continue;
00199 }
00200
00201 after_header = std::chrono::steady_clock::now();
00202
00203 if (Fragment::isUserFragmentType(header.type) || header.type == Fragment::DataFragmentType || header.type == Fragment::EmptyFragmentType || header.type == Fragment::ContainerFragmentType) {
00204 TLOG(TLVL_TRACE) << "Received Fragment Header from rank " << source_rank << ".";
00205 RawDataType* loc = nullptr;
00206 size_t retries = 0;
00207 while (loc == nullptr)
00208 {
00209 loc = shm_manager_->WriteFragmentHeader(header);
00210 if (loc == nullptr) usleep(sleep_time);
00211 if (stop_requested_) return;
00212 retries++;
00213 if (non_reliable_mode_enabled_ && retries > max_retries)
00214 {
00215 loc = shm_manager_->WriteFragmentHeader(header, true);
00216 }
00217 }
00218 if (loc == nullptr)
00219 {
00220
00221 TLOG(TLVL_ERROR) << "runReceiver_: Could not get data location for event " << header.sequence_id;
00222 continue;
00223 }
00224 before_body = std::chrono::steady_clock::now();
00225
00226 TLOG(16) << "runReceiver_: Calling receiveFragmentData";
00227 auto ret2 = source_plugins_[source_rank]->receiveFragmentData(loc, header.word_count - header.num_words());
00228 TLOG(16) << "runReceiver_: Done with receiveFragmentData, ret2=" << ret2 << " (should be " << source_rank << ")";
00229
00230 if (ret != ret2) {
00231 TLOG(TLVL_ERROR) << "Unexpected return code from receiveFragmentData after receiveFragmentHeader! (Expected: " << ret << ", Got: " << ret2 << ")";
00232 throw cet::exception("DataReceiverManager") << "Unexpected return code from receiveFragmentData after receiveFragmentHeader! (Expected: " << ret << ", Got: " << ret2 << ")";
00233 }
00234
00235 shm_manager_->DoneWritingFragment(header);
00236 TLOG(TLVL_TRACE) << "Done receiving fragment with sequence ID " << header.sequence_id << " from rank " << source_rank;
00237
00238 recv_frag_count_.incSlot(source_rank);
00239 recv_frag_size_.incSlot(source_rank, header.word_count * sizeof(RawDataType));
00240 recv_seq_count_.setSlot(source_rank, header.sequence_id);
00241 if (endOfDataCount != static_cast<size_t>(-1))
00242 {
00243 TLOG(TLVL_DEBUG) << "Received fragment " << header.sequence_id << " from rank " << source_rank
00244 << " (" << recv_frag_count_.slotCount(source_rank) << "/" << endOfDataCount << ")";
00245 }
00246
00247
00248 source_metric_data_[source_rank].delta_t += TimeUtils::GetElapsedTime(start_time);
00249 source_metric_data_[source_rank].hdr_delta_t += TimeUtils::GetElapsedTime(start_time, after_header);
00250 source_metric_data_[source_rank].store_delta_t += TimeUtils::GetElapsedTime(after_header, before_body);
00251 source_metric_data_[source_rank].data_delta_t += TimeUtils::GetElapsedTime(before_body);
00252
00253 source_metric_data_[source_rank].data_size += header.word_count * sizeof(RawDataType);
00254 source_metric_data_[source_rank].header_size += header.num_words() * sizeof(RawDataType);
00255
00256 if (metricMan && TimeUtils::GetElapsedTime(source_metric_send_time_[source_rank]) > 1)
00257 {
00258 TLOG(6) << "runReceiver_: Sending receive stats";
00259 metricMan->sendMetric("Total Receive Time From Rank " + std::to_string(source_rank), source_metric_data_[source_rank].delta_t, "s", 5, MetricMode::Accumulate);
00260 metricMan->sendMetric("Total Receive Size From Rank " + std::to_string(source_rank), static_cast<unsigned long>(source_metric_data_[source_rank].data_size), "B", 5, MetricMode::Accumulate);
00261 metricMan->sendMetric("Total Receive Rate From Rank " + std::to_string(source_rank), source_metric_data_[source_rank].data_size / source_metric_data_[source_rank].delta_t, "B/s", 5, MetricMode::Average);
00262
00263 metricMan->sendMetric("Header Receive Time From Rank " + std::to_string(source_rank), source_metric_data_[source_rank].hdr_delta_t, "s", 5, MetricMode::Accumulate);
00264 metricMan->sendMetric("Header Receive Size From Rank " + std::to_string(source_rank), static_cast<unsigned long>(source_metric_data_[source_rank].header_size), "B", 5, MetricMode::Accumulate);
00265 metricMan->sendMetric("Header Receive Rate From Rank " + std::to_string(source_rank), source_metric_data_[source_rank].header_size / source_metric_data_[source_rank].hdr_delta_t, "B/s", 5, MetricMode::Average);
00266
00267 auto payloadSize = source_metric_data_[source_rank].data_size - source_metric_data_[source_rank].header_size;
00268 metricMan->sendMetric("Data Receive Time From Rank " + std::to_string(source_rank), source_metric_data_[source_rank].data_delta_t, "s", 5, MetricMode::Accumulate);
00269 metricMan->sendMetric("Data Receive Size From Rank " + std::to_string(source_rank), static_cast<unsigned long>(payloadSize), "B", 5, MetricMode::Accumulate);
00270 metricMan->sendMetric("Data Receive Rate From Rank " + std::to_string(source_rank), payloadSize / source_metric_data_[source_rank].data_delta_t, "B/s", 5, MetricMode::Average);
00271
00272 metricMan->sendMetric("Data Receive Count From Rank " + std::to_string(source_rank), recv_frag_count_.slotCount(source_rank), "fragments", 3, MetricMode::LastPoint);
00273
00274 metricMan->sendMetric("Shared Memory Wait Time From Rank " + std::to_string(source_rank), source_metric_data_[source_rank].store_delta_t, "s", 5, MetricMode::Accumulate);
00275
00276 TLOG(6) << "runReceiver_: Done sending receive stats";
00277
00278 source_metric_send_time_[source_rank] = std::chrono::steady_clock::now();
00279 source_metric_data_[source_rank] = source_metric_data();
00280 }
00281 }
00282 else if (header.type == Fragment::EndOfDataFragmentType || header.type == Fragment::InitFragmentType || header.type == Fragment::EndOfRunFragmentType || header.type == Fragment::EndOfSubrunFragmentType || header.type == Fragment::ShutdownFragmentType)
00283 {
00284 TLOG(TLVL_DEBUG) << "Received System Fragment from rank " << source_rank << " of type " << detail::RawFragmentHeader::SystemTypeToString(header.type) << ".";
00285
00286 FragmentPtr frag(new Fragment(header.word_count - header.num_words()));
00287 memcpy(frag->headerAddress(), &header, header.num_words() * sizeof(RawDataType));
00288 auto ret3 = source_plugins_[source_rank]->receiveFragmentData(frag->headerAddress() + header.num_words(), header.word_count - header.num_words());
00289 if (ret3 != source_rank)
00290 {
00291 TLOG(TLVL_ERROR) << "Unexpected return code from receiveFragmentData after receiveFragmentHeader while receiving System Fragment! (Expected: " << source_rank << ", Got: " << ret3 << ")";
00292 throw cet::exception("DataReceiverManager") << "Unexpected return code from receiveFragmentData after receiveFragmentHeader while receiving System Fragment! (Expected: " << source_rank << ", Got: " << ret3 << ")";
00293 }
00294
00295 switch (header.type)
00296 {
00297 case Fragment::EndOfDataFragmentType:
00298 shm_manager_->setRequestMode(detail::RequestMessageMode::EndOfRun);
00299 endOfDataCount = *(frag->dataBegin());
00300 TLOG(TLVL_DEBUG) << "EndOfData Fragment indicates that " << endOfDataCount << " fragments are expected from rank " << source_rank
00301 << " (recvd " << recv_frag_count_.slotCount(source_rank) << ").";
00302 break;
00303 case Fragment::InitFragmentType:
00304 TLOG(TLVL_DEBUG) << "Received Init Fragment from rank " << source_rank << ".";
00305 shm_manager_->setRequestMode(detail::RequestMessageMode::Normal);
00306 shm_manager_->SetInitFragment(std::move(frag));
00307 break;
00308 case Fragment::EndOfRunFragmentType:
00309 shm_manager_->setRequestMode(detail::RequestMessageMode::EndOfRun);
00310
00311 break;
00312 case Fragment::EndOfSubrunFragmentType:
00313
00314 if (header.sequence_id != Fragment::InvalidSequenceID) shm_manager_->rolloverSubrun(header.sequence_id);
00315 else shm_manager_->rolloverSubrun(recv_seq_count_.slotCount(source_rank));
00316 break;
00317 case Fragment::ShutdownFragmentType:
00318 shm_manager_->setRequestMode(detail::RequestMessageMode::EndOfRun);
00319 break;
00320 }
00321 }
00322 }
00323
00324
00325 running_sources_[source_rank] = false;
00326 }