TritonService Class Reference

#include <TritonService.h>

Classes
struct	FallbackOpts
struct	Model
struct	Module
struct	Server

Public Member Functions
void	addModel (const std::string &modelName, const std::string &path)
void	notifyCallStatus (bool status) const
const std::string &	pid () const
Server	serverInfo (const std::string &model, const std::string &preferred="") const
	TritonService (const edm::ParameterSet &pset, edm::ActivityRegistry &areg)
	~TritonService ()=default

Static Public Member Functions
static void	fillDescriptions (edm::ConfigurationDescriptions &descriptions)

Private Member Functions
void	postEndJob ()
void	postModuleConstruction (edm::ModuleDescription const &)
void	preallocate (edm::service::SystemBounds const &)
void	preBeginJob (edm::PathsAndConsumesOfModulesBase const &, edm::ProcessContext const &)
void	preModuleConstruction (edm::ModuleDescription const &)
void	preModuleDestruction (edm::ModuleDescription const &)
template<typename LOG >
void	printFallbackServerLog () const

Private Attributes
bool	allowAddModel_
std::atomic< int >	callFails_
unsigned	currentModuleId_
FallbackOpts	fallbackOpts_
std::unordered_map< std::string, Model >	models_
std::unordered_map< unsigned, Module >	modules_
int	numberOfThreads_
std::string	pid_
std::unordered_map< std::string, Server >	servers_
bool	startedFallback_
std::unordered_map< std::string, Model >	unservedModels_
bool	verbose_

Detailed Description

Definition at line 31 of file TritonService.h.

Constructor & Destructor Documentation

TritonService()

TritonService::TritonService	(	const edm::ParameterSet &	pset,
		edm::ActivityRegistry &	areg
	)

Definition at line 58 of file TritonService.cc.

References HLTObjectMonitor_Client_cff::client, TritonService::FallbackOpts::enable, Exception, TritonService::Server::fallbackAddress, TritonService::Server::fallbackName, fallbackOpts_, LocalCPU, LocalGPU, ecalBarrelClusterFastTimer_cfi::modelName, models_, mps_check::msg, postEndJob(), postModuleConstruction(), preallocate(), preBeginJob(), preModuleConstruction(), preModuleDestruction(), muonDTDigis_cfi::pset, Remote, contentValuesFiles::server, servers_, AlCaHLTBitMon_QueryRunRegistry::string, TRITON_THROW_IF_ERROR, tier0::unique(), TritonService::FallbackOpts::useGPU, verbose_, edm::ActivityRegistry::watchPostEndJob(), edm::ActivityRegistry::watchPostModuleConstruction(), edm::ActivityRegistry::watchPreallocate(), edm::ActivityRegistry::watchPreBeginJob(), edm::ActivityRegistry::watchPreModuleConstruction(), and edm::ActivityRegistry::watchPreModuleDestruction().

    : verbose_(pset.getUntrackedParameter<bool>("verbose")),
      fallbackOpts_(pset.getParameterSet("fallback")),
      currentModuleId_(0),
      allowAddModel_(false),
      startedFallback_(false),
      callFails_(0),
      pid_(std::to_string(::getpid())) {
  //module construction is assumed to be serial (correct at the time this code was written)

  areg.watchPreallocate(this, &TritonService::preallocate);

  areg.watchPreModuleConstruction(this, &TritonService::preModuleConstruction);
  areg.watchPostModuleConstruction(this, &TritonService::postModuleConstruction);
  areg.watchPreModuleDestruction(this, &TritonService::preModuleDestruction);
  //fallback server will be launched (if needed) before beginJob
  areg.watchPreBeginJob(this, &TritonService::preBeginJob);
  areg.watchPostEndJob(this, &TritonService::postEndJob);

  //include fallback server in set if enabled
  if (fallbackOpts_.enable) {
    auto serverType = TritonServerType::Remote;
    if (!fallbackOpts_.useGPU)
      serverType = TritonServerType::LocalCPU;
#ifdef TRITON_ENABLE_GPU
    else
      serverType = TritonServerType::LocalGPU;
#endif

    servers_.emplace(std::piecewise_construct,
                     std::forward_as_tuple(Server::fallbackName),
                     std::forward_as_tuple(Server::fallbackName, Server::fallbackAddress, serverType));
  }

  //loop over input servers: check which models they have
  std::string msg;
  if (verbose_)
    msg = "List of models for each server:\n";
  for (const auto& serverPset : pset.getUntrackedParameterSetVector("servers")) {
    const std::string& serverName(serverPset.getUntrackedParameter<std::string>("name"));
    //ensure uniqueness
    auto [sit, unique] = servers_.emplace(serverName, serverPset);
    if (!unique)
      throw cms::Exception("DuplicateServer")
          << "TritonService: Not allowed to specify more than one server with same name (" << serverName << ")";
    auto& server(sit->second);

    std::unique_ptr<tc::InferenceServerGrpcClient> client;
    TRITON_THROW_IF_ERROR(
        tc::InferenceServerGrpcClient::Create(&client, server.url, false, server.useSsl, server.sslOptions),
        "TritonService(): unable to create inference context for " + serverName + " (" + server.url + ")",
        false);

    if (verbose_) {
      inference::ServerMetadataResponse serverMetaResponse;
      TRITON_THROW_IF_ERROR(client->ServerMetadata(&serverMetaResponse),
                            "TritonService(): unable to get metadata for " + serverName + " (" + server.url + ")",
                            false);
      edm::LogInfo("TritonService") << "Server " << serverName << ": url = " << server.url
                                    << ", version = " << serverMetaResponse.version();
    }

    inference::RepositoryIndexResponse repoIndexResponse;
    TRITON_THROW_IF_ERROR(client->ModelRepositoryIndex(&repoIndexResponse),
                          "TritonService(): unable to get repository index for " + serverName + " (" + server.url + ")",
                          false);

    //servers keep track of models and vice versa
    if (verbose_)
      msg += serverName + ": ";
    for (const auto& modelIndex : repoIndexResponse.models()) {
      const auto& modelName = modelIndex.name();
      auto mit = models_.find(modelName);
      if (mit == models_.end())
        mit = models_.emplace(modelName, "").first;
      auto& modelInfo(mit->second);
      modelInfo.servers.insert(serverName);
      server.models.insert(modelName);
      if (verbose_)
        msg += modelName + ", ";
    }
    if (verbose_)
      msg += "\n";
  }
  if (verbose_)
    edm::LogInfo("TritonService") << msg;
}

~TritonService()

TritonService::~TritonService ( )

default

Member Function Documentation

addModel()

void TritonService::addModel	(	const std::string &	modelName,
		const std::string &	path
	)

Definition at line 155 of file TritonService.cc.

References allowAddModel_, currentModuleId_, Exception, ecalBarrelClusterFastTimer_cfi::modelName, models_, modules_, castor_dqm_sourceclient_file_cfg::path, and unservedModels_.

Referenced by TritonEDFilterT< G, Capabilities >::initializeGlobalCache(), TritonEDProducerT< G, Capabilities >::initializeGlobalCache(), and TritonOneEDAnalyzer< Capabilities >::TritonOneEDAnalyzer().

                                                                              {
  //should only be called in module constructors
  if (!allowAddModel_)
    throw cms::Exception("DisallowedAddModel")
        << "TritonService: Attempt to call addModel() outside of module constructors";
  //if model is not in the list, then no specified server provides it
  auto mit = models_.find(modelName);
  if (mit == models_.end()) {
    auto& modelInfo(unservedModels_.emplace(modelName, path).first->second);
    modelInfo.modules.insert(currentModuleId_);
    //only keep track of modules that need unserved models
    modules_.emplace(currentModuleId_, modelName);
  }
}

fillDescriptions()

void TritonService::fillDescriptions ( edm::ConfigurationDescriptions &  descriptions )

static

Definition at line 362 of file TritonService.cc.

References edm::ParameterSetDescription::addUntracked(), edm::ConfigurationDescriptions::addWithDefaultLabel(), submitPVResolutionJobs::desc, and AlCaHLTBitMon_QueryRunRegistry::string.

                                                                               {
  edm::ParameterSetDescription desc;
  desc.addUntracked<bool>("verbose", false);

  edm::ParameterSetDescription validator;
  validator.addUntracked<std::string>("name");
  validator.addUntracked<std::string>("address");
  validator.addUntracked<unsigned>("port");
  validator.addUntracked<bool>("useSsl", false);
  validator.addUntracked<std::string>("rootCertificates", "");
  validator.addUntracked<std::string>("privateKey", "");
  validator.addUntracked<std::string>("certificateChain", "");

  desc.addVPSetUntracked("servers", validator, {});

  edm::ParameterSetDescription fallbackDesc;
  fallbackDesc.addUntracked<bool>("enable", false);
  fallbackDesc.addUntracked<bool>("debug", false);
  fallbackDesc.addUntracked<bool>("verbose", false);
  fallbackDesc.addUntracked<bool>("useDocker", false);
  fallbackDesc.addUntracked<bool>("useGPU", false);
  fallbackDesc.addUntracked<int>("retries", -1);
  fallbackDesc.addUntracked<int>("wait", -1);
  fallbackDesc.addUntracked<std::string>("instanceBaseName", "triton_server_instance");
  fallbackDesc.addUntracked<std::string>("instanceName", "");
  fallbackDesc.addUntracked<std::string>("tempDir", "");
  fallbackDesc.addUntracked<std::string>("imageName", "");
  fallbackDesc.addUntracked<std::string>("sandboxName", "");
  desc.add<edm::ParameterSetDescription>("fallback", fallbackDesc);

  descriptions.addWithDefaultLabel(desc);
}

notifyCallStatus()

void TritonService::notifyCallStatus

(

bool

status

)

const

Definition at line 303 of file TritonService.cc.

References callFails_, and mps_update::status.

Referenced by TritonClient::evaluate(), and TritonException::TritonException().

                                                      {
  if (status)
    --callFails_;
  else
    ++callFails_;
}

pid()

const std::string& TritonService::pid ( ) const

inline

Definition at line 115 of file TritonService.h.

References pid_.

Referenced by TritonClient::TritonClient().

{ return pid_; }

postEndJob()

void TritonService::postEndJob ( )

private

Definition at line 310 of file TritonService.cc.

References callFails_, mps_check::command, TritonService::FallbackOpts::command, Exception, fallbackOpts_, dqm-mbProfile::format, or, convertSQLitetoXML_cfg::output, startedFallback_, AlCaHLTBitMon_QueryRunRegistry::string, and verbose_.

Referenced by TritonService().

                               {
  if (!startedFallback_)
    return;

  std::string command = fallbackOpts_.command;
  //prevent log cleanup during server stop
  if (callFails_ > 0)
    command += " -c";
  command += " stop";
  if (verbose_)
    edm::LogInfo("TritonService") << command;

  const auto& [output, rv] = execSys(command);
  if (rv != 0 or callFails_ > 0) {
    //print logs if cmsRun is currently exiting because of a TritonException
    edm::LogError("TritonService") << output;
    printFallbackServerLog<edm::LogError>();
    if (rv != 0) {
      std::string stopCat("FallbackFailed");
      std::string stopMsg = fmt::format("TritonService: Stopping the fallback server failed with exit code {}", rv);
      //avoid throwing if the stack is already unwinding
      if (callFails_ > 0)
        edm::LogWarning(stopCat) << stopMsg;
      else
        throw cms::Exception(stopCat) << stopMsg;
    }
  } else if (verbose_) {
    edm::LogInfo("TritonService") << output;
    printFallbackServerLog<edm::LogInfo>();
  }
}

postModuleConstruction()

void TritonService::postModuleConstruction ( edm::ModuleDescription const &  desc )

private

Definition at line 170 of file TritonService.cc.

References allowAddModel_.

Referenced by TritonService().

{ allowAddModel_ = false; }

preallocate()

void TritonService::preallocate ( edm::service::SystemBounds const &  bounds )

private

Definition at line 146 of file TritonService.cc.

References numberOfThreads_.

Referenced by TritonService().

                                                                    {
  numberOfThreads_ = bounds.maxNumberOfThreads();
}

preBeginJob()

void TritonService::preBeginJob ( edm::PathsAndConsumesOfModulesBase const &  , edm::ProcessContext const &    )

private

Definition at line 215 of file TritonService.cc.

References mps_check::command, TritonService::FallbackOpts::command, edm::createGlobalIdentifier(), TritonService::FallbackOpts::debug, TritonService::FallbackOpts::enable, Exception, TritonService::Server::fallbackName, fallbackOpts_, ALPAKA_ACCELERATOR_NAMESPACE::caPixelDoublets::if(), TritonService::FallbackOpts::imageName, TritonService::FallbackOpts::instanceName, ReggeGribovPartonMC_EposLHC_2760GeV_PbPb_cfi::model, ecalBarrelClusterFastTimer_cfi::modelName, models_, mps_check::msg, numberOfThreads_, or, convertSQLitetoXML_cfg::output, pid_, TritonService::FallbackOpts::retries, TritonService::FallbackOpts::sandboxName, contentValuesFiles::server, servers_, startedFallback_, AlCaHLTBitMon_QueryRunRegistry::string, TritonService::FallbackOpts::tempDir, to_string(), unservedModels_, TritonService::FallbackOpts::useDocker, TritonService::FallbackOpts::useGPU, TritonService::FallbackOpts::verbose, verbose_, and TritonService::FallbackOpts::wait.

Referenced by TritonService().

                                                                                                 {
  //only need fallback if there are unserved models
  if (!fallbackOpts_.enable or unservedModels_.empty())
    return;

  std::string msg;
  if (verbose_)
    msg = "List of models for fallback server: ";
  //all unserved models are provided by fallback server
  auto& server(servers_.find(Server::fallbackName)->second);
  for (const auto& [modelName, model] : unservedModels_) {
    auto& modelInfo(models_.emplace(modelName, model).first->second);
    modelInfo.servers.insert(Server::fallbackName);
    server.models.insert(modelName);
    if (verbose_)
      msg += modelName + ", ";
  }
  if (verbose_)
    edm::LogInfo("TritonService") << msg;

  //assemble server start command
  fallbackOpts_.command = "cmsTriton -P -1 -p " + pid_;
  if (fallbackOpts_.debug)
    fallbackOpts_.command += " -c";
  if (fallbackOpts_.verbose)
    fallbackOpts_.command += " -v";
  if (fallbackOpts_.useDocker)
    fallbackOpts_.command += " -d";
  if (fallbackOpts_.useGPU)
    fallbackOpts_.command += " -g";
  if (!fallbackOpts_.instanceName.empty())
    fallbackOpts_.command += " -n " + fallbackOpts_.instanceName;
  if (fallbackOpts_.retries >= 0)
    fallbackOpts_.command += " -r " + std::to_string(fallbackOpts_.retries);
  if (fallbackOpts_.wait >= 0)
    fallbackOpts_.command += " -w " + std::to_string(fallbackOpts_.wait);
  for (const auto& [modelName, model] : unservedModels_) {
    fallbackOpts_.command += " -m " + model.path;
  }
  std::string thread_string = " -I " + std::to_string(numberOfThreads_);
  fallbackOpts_.command += thread_string;
  if (!fallbackOpts_.imageName.empty())
    fallbackOpts_.command += " -i " + fallbackOpts_.imageName;
  if (!fallbackOpts_.sandboxName.empty())
    fallbackOpts_.command += " -s " + fallbackOpts_.sandboxName;
  //don't need this anymore
  unservedModels_.clear();

  //get a random temporary directory if none specified
  if (fallbackOpts_.tempDir.empty()) {
    auto tmp_dir_path{std::filesystem::temp_directory_path() /= edm::createGlobalIdentifier()};
    fallbackOpts_.tempDir = tmp_dir_path.string();
  }
  //special case ".": use script default (temp dir = .$instanceName)
  if (fallbackOpts_.tempDir != ".")
    fallbackOpts_.command += " -t " + fallbackOpts_.tempDir;

  std::string command = fallbackOpts_.command + " start";

  if (fallbackOpts_.debug)
    edm::LogInfo("TritonService") << "Fallback server temporary directory: " << fallbackOpts_.tempDir;
  if (verbose_)
    edm::LogInfo("TritonService") << command;

  //mark as started before executing in case of ctrl+c while command is running
  startedFallback_ = true;
  const auto& [output, rv] = execSys(command);
  if (rv != 0) {
    edm::LogError("TritonService") << output;
    printFallbackServerLog<edm::LogError>();
    throw cms::Exception("FallbackFailed")
        << "TritonService: Starting the fallback server failed with exit code " << rv;
  } else if (verbose_)
    edm::LogInfo("TritonService") << output;
  //get the port
  const std::string& portIndicator("CMS_TRITON_GRPC_PORT: ");
  //find last instance in log in case multiple ports were tried
  auto pos = output.rfind(portIndicator);
  if (pos != std::string::npos) {
    auto pos2 = pos + portIndicator.size();
    auto pos3 = output.find('\n', pos2);
    const auto& portNum = output.substr(pos2, pos3 - pos2);
    server.url += ":" + portNum;
  } else
    throw cms::Exception("FallbackFailed") << "TritonService: Unknown port for fallback server, log follows:\n"
                                           << output;
}

preModuleConstruction()

void TritonService::preModuleConstruction ( edm::ModuleDescription const &  desc )

private

Definition at line 150 of file TritonService.cc.

References allowAddModel_, currentModuleId_, and submitPVResolutionJobs::desc.

Referenced by TritonService().

                                                                          {
  currentModuleId_ = desc.id();
  allowAddModel_ = true;
}

preModuleDestruction()

void TritonService::preModuleDestruction ( edm::ModuleDescription const &  desc )

private

Definition at line 172 of file TritonService.cc.

References submitPVResolutionJobs::desc, modules_, and unservedModels_.

Referenced by TritonService().

                                                                         {
  //remove destructed modules from unserved list
  if (unservedModels_.empty())
    return;
  auto id = desc.id();
  auto oit = modules_.find(id);
  if (oit != modules_.end()) {
    const auto& moduleInfo(oit->second);
    auto mit = unservedModels_.find(moduleInfo.model);
    if (mit != unservedModels_.end()) {
      auto& modelInfo(mit->second);
      modelInfo.modules.erase(id);
      //remove a model if it is no longer needed by any modules
      if (modelInfo.modules.empty())
        unservedModels_.erase(mit);
    }
    modules_.erase(oit);
  }
}

printFallbackServerLog()

template<typename LOG >

void TritonService::printFallbackServerLog ( ) const

private

Definition at line 343 of file TritonService.cc.

References fallbackOpts_, timingPdfMaker::infile, TritonService::FallbackOpts::instanceName, LOG, heppy_check::logName, and TritonService::FallbackOpts::tempDir.

                                                 {
  std::vector<std::string> logNames{"log_" + fallbackOpts_.instanceName + ".log"};
  //cmsTriton script moves log from temp to current dir in verbose mode or in some cases when auto_stop is called
  // -> check both places
  logNames.push_back(fallbackOpts_.tempDir + "/" + logNames[0]);
  bool foundLog = false;
  for (const auto& logName : logNames) {
    std::ifstream infile(logName);
    if (infile.is_open()) {
      LOG("TritonService") << "TritonService: server log " << logName << "\n" << infile.rdbuf();
      foundLog = true;
      break;
    }
  }
  if (!foundLog)
    LOG("TritonService") << "TritonService: could not find server log " << logNames[0] << " in current directory or "
                         << fallbackOpts_.tempDir;
}

serverInfo()

TritonService::Server TritonService::serverInfo	(	const std::string &	model,
		const std::string &	preferred = ""
	)		const

Definition at line 193 of file TritonService.cc.

References ReggeGribovPartonMC_EposLHC_2760GeV_PbPb_cfi::model, models_, contentValuesFiles::server, and servers_.

Referenced by TritonClient::TritonClient().

                                                                                                      {
  auto mit = models_.find(model);
  if (mit == models_.end())
    throw cms::Exception("MissingModel") << "TritonService: There are no servers that provide model " << model;
  const auto& modelInfo(mit->second);
  const auto& modelServers = modelInfo.servers;

  auto msit = modelServers.end();
  if (!preferred.empty()) {
    msit = modelServers.find(preferred);
    //todo: add a "strict" parameter to stop execution if preferred server isn't found?
    if (msit == modelServers.end())
      edm::LogWarning("PreferredServer") << "Preferred server " << preferred << " for model " << model
                                         << " not available, will choose another server";
  }
  const auto& serverName(msit == modelServers.end() ? *modelServers.begin() : preferred);

  //todo: use some algorithm to select server rather than just picking arbitrarily
  const auto& server(servers_.find(serverName)->second);
  return server;
}

Member Data Documentation

allowAddModel_

bool TritonService::allowAddModel_

private

Definition at line 135 of file TritonService.h.

Referenced by addModel(), postModuleConstruction(), and preModuleConstruction().

callFails_

std::atomic<int> TritonService::callFails_

mutableprivate

Definition at line 137 of file TritonService.h.

Referenced by notifyCallStatus(), and postEndJob().

currentModuleId_

unsigned TritonService::currentModuleId_

private

Definition at line 134 of file TritonService.h.

Referenced by addModel(), and preModuleConstruction().

fallbackOpts_

FallbackOpts TritonService::fallbackOpts_

private

Definition at line 133 of file TritonService.h.

Referenced by postEndJob(), preBeginJob(), printFallbackServerLog(), and TritonService().

models_

std::unordered_map<std::string, Model> TritonService::models_

private

Definition at line 142 of file TritonService.h.

Referenced by addModel(), preBeginJob(), serverInfo(), and TritonService().

modules_

std::unordered_map<unsigned, Module> TritonService::modules_

private

Definition at line 143 of file TritonService.h.

Referenced by addModel(), and preModuleDestruction().

numberOfThreads_

int TritonService::numberOfThreads_

private

Definition at line 144 of file TritonService.h.

Referenced by preallocate(), and preBeginJob().

pid_

std::string TritonService::pid_

private

Definition at line 138 of file TritonService.h.

Referenced by pid(), and preBeginJob().

servers_

std::unordered_map<std::string, Server> TritonService::servers_

private

Definition at line 141 of file TritonService.h.

Referenced by preBeginJob(), serverInfo(), and TritonService().

startedFallback_

bool TritonService::startedFallback_

private

Definition at line 136 of file TritonService.h.

Referenced by postEndJob(), and preBeginJob().

unservedModels_

std::unordered_map<std::string, Model> TritonService::unservedModels_

private

Definition at line 139 of file TritonService.h.

Referenced by addModel(), preBeginJob(), and preModuleDestruction().

verbose_

bool TritonService::verbose_

private

Definition at line 132 of file TritonService.h.

Referenced by postEndJob(), preBeginJob(), and TritonService().