BatchDriver.cxx File Reference

#include <EventLoop/BatchDriver.h>
#include <EventLoop/BatchJob.h>
#include <EventLoop/BatchSample.h>
#include <EventLoop/BatchSegment.h>
#include <EventLoop/ManagerData.h>
#include <EventLoop/ManagerStep.h>
#include <EventLoop/MessageCheck.h>
#include <EventLoop/OutputStream.h>
#include <RootCoreUtils/Assert.h>
#include <RootCoreUtils/ShellExec.h>
#include <RootCoreUtils/StringUtil.h>
#include <RootCoreUtils/hadd.h>
#include <SampleHandler/DiskOutputLocal.h>
#include <SampleHandler/MetaFields.h>
#include <SampleHandler/MetaVector.h>
#include <SampleHandler/Sample.h>
#include <TFile.h>
#include <TSystem.h>
#include <cmath>
#include <cstdlib>
#include <fstream>
#include <memory>
#include <sstream>
#include <stdexcept>

Go to the source code of this file.

Functions
	ClassImp (EL::BatchDriver) namespace EL

Function Documentation

ClassImp()

ClassImp

(

EL::BatchDriver

)

Author

Nils Krumnack

effects: fill the job object from the original job information except for the actual samples submitted guarantee: basic failures: out of memory II

effects: fill the given sample object with the information from the SampleHandler object, except for the segments that get filled separately guarantee: basic failures: out of memory II

effects: add the given sample with the given segments to the job guarantee: basic failures: out of memory II failures: segment misconfiguration requires: !segments.empty()

effects: generate the splits for the sample by file guarantee: basic failures: out of memory II failures: i/o errors requires: !sample.files.empty() requires: numJobs == rint (numJobs) postcondition: !segments.empty()

effects: generate the splits for the sample guarantee: basic failures: out of memory II failures: i/o errors requires: !sample.files.empty() requires: numJobs == rint (numJobs) postcondition: !segments.empty()

effects: generate the splits for the sample guarantee: basic failures: out of memory II failures: i/o errors requires: !sample.files.empty() postcondition: !segments.empty()

effects: fill the job object from the original job information including the actual samples submitted guarantee: basic failures: out of memory II

Definition at line 42 of file BatchDriver.cxx.

{
  namespace
  {
    void fillJob (BatchJob& myjob, const Job& job,
                  const std::string& submitDir)
    {
      myjob.job = job;
      myjob.location = submitDir;
    }
 
 

    void fillSample (BatchSample& mysample,
                     const SH::Sample& sample, const SH::MetaObject& meta)
    {
      mysample.name = sample.name();
      mysample.meta = *sample.meta();
      mysample.meta.fetchDefaults (meta);
 
      mysample.files = sample.makeFileList ();
    }
 
 

    void addSample (BatchJob& job, BatchSample sample,
                    const std::vector<BatchSegment>& segments)
    {
      RCU_REQUIRE (!segments.empty());
 
      sample.begin_segments = job.segments.size();
 
      RCU_ASSERT (segments[0].begin_file == 0);
      RCU_ASSERT (segments[0].begin_event == 0);
      for (std::size_t iter = 0, end = segments.size(); iter != end; ++ iter)
      {
        BatchSegment segment = segments[iter];
 
        segment.sampleName = sample.name;
        {
          std::ostringstream myname;
          myname << sample.name << "-" << iter;
          segment.fullName = myname.str();
        }
        {
          std::ostringstream myname;
          myname << iter;
          segment.segmentName = myname.str();
        }
 
        segment.sample = job.samples.size();
        segment.job_id = job.segments.size();
        if (iter+1 < end)
        {
          segment.end_file  = segments[iter+1].begin_file;
          segment.end_event = segments[iter+1].begin_event;
        } else
        {
          segment.end_file = sample.files.size();
          segment.end_event = 0;
        }
        RCU_ASSERT (segment.begin_file < segment.end_file || (segment.begin_file == segment.end_file && segment.begin_event <= segment.end_event));
        job.segments.push_back (segment);
      }
 
      sample.end_segments = job.segments.size();
      job.samples.push_back (sample);
    }
 
 
 
 

    void splitSampleByFile (const BatchSample& sample,
                            std::vector<BatchSegment>& segments,
                            double numJobs)
    {
      // RCU_REQUIRE (!sample.files.empty());
      RCU_REQUIRE (numJobs == rint (numJobs));
      RCU_REQUIRE (numJobs <= sample.files.size());
 
      for (std::size_t index = 0, end = numJobs;
           index < end; ++ index)
      {
        BatchSegment segment;
        segment.begin_file = rint (index * (sample.files.size() / numJobs));
        segments.push_back (segment);
      }
    }
 
 
 
 

    void splitSampleByEvent (const BatchSample& sample,
                             std::vector<BatchSegment>& segments,
                             const std::vector<Long64_t>& eventsFile,
                             double numJobs)
    {
      RCU_REQUIRE (!sample.files.empty());
      RCU_REQUIRE (sample.files.size() == eventsFile.size());
      RCU_REQUIRE (numJobs == rint (numJobs));
 
      Long64_t eventsSum = 0;
      for (std::vector<Long64_t>::const_iterator nevents = eventsFile.begin(),
             end = eventsFile.end(); nevents != end; ++ nevents)
      {
        RCU_ASSERT_SOFT (*nevents >= 0);
        eventsSum += *nevents;
      }
      if (eventsSum == 0)
      {
        // an all-empty sample still needs a single segment so that its
        // (empty) files get processed, e.g. for metadata; there is no
        // per-event split to compute and the division by numJobs below
        // would be undefined for numJobs == 0.
        segments.push_back (BatchSegment ());
        return;
      }
      if (numJobs > eventsSum)
        numJobs = eventsSum;
      Long64_t eventsMax
        = sample.meta.castDouble (Job::optEventsPerWorker);
      if (eventsMax > 0)
        numJobs = ceil (double (eventsSum) / eventsMax);
      eventsMax = Long64_t (ceil (double (eventsSum) / numJobs));
 
      BatchSegment segment;
      while (std::size_t (segment.begin_file) < eventsFile.size())
      {
        while (segment.begin_event < eventsFile[segment.begin_file])
        {
          segments.push_back (segment);
          segment.begin_event += eventsMax;
        }
        segment.begin_event -= eventsFile[segment.begin_file];
        ++ segment.begin_file;
      }
      // eventsMax is rounded up, so the number of segments produced can be
      // fewer than the requested number of jobs (never more) for inputs
      // that do not divide evenly.
      RCU_ASSERT (segments.size() <= numJobs);
    }
 
 
 
 

    void splitSample (const BatchSample& sample,
                      std::vector<BatchSegment>& segments)
    {
      // RCU_REQUIRE (!sample.files.empty());
 
      // rationale: I'm now calculating the number of jobs here, so
      //   that independent of whether we divide on a per-file or
      //   per-event basis we can get the same number of jobs.  also,
      //   it will balance out things slightly if we only have a few
      //   files and multiple files per job.
      const double filesPerWorker
        = sample.meta.castDouble (Job::optFilesPerWorker, 1);
      if (filesPerWorker < 1)
      {
        std::ostringstream msg;
        msg << "invalid number of files per worker: " << filesPerWorker;
        throw std::runtime_error (msg.str());
      }
      double numJobs = ceil (sample.files.size() / filesPerWorker);
 
      const TObject *meta
        = sample.meta.get (SH::MetaFields::numEventsPerFile);
      if (meta)
      {
        const SH::MetaVector<Long64_t> *const meta_nentries
          = dynamic_cast<const SH::MetaVector<Long64_t> *>(meta);
        RCU_ASSERT_SOFT (meta == meta_nentries);
        RCU_ASSERT_SOFT (meta_nentries->value.size() == sample.files.size());
        splitSampleByEvent (sample, segments, meta_nentries->value, numJobs);
      } else
      {
        splitSampleByFile (sample, segments, numJobs);
      }
 
 
      if (segments.empty())
      {
        // rationale: this isn't really the proper thing to do.  if a
        //   sample is empty I should just run the job locally.
        BatchSegment empty;
        segments.push_back (empty);
      }
 
      RCU_PROVIDE (!segments.empty());
    }
 

    void fillFullJob (BatchJob& myjob, const Job& job,
                      const std::string& location,
                      const SH::MetaObject& meta)
    {
      fillJob (myjob, job, location);
      *myjob.job.options() = meta;
 
      for (const SH::Sample *sample : job.sampleHandler())
      {
        BatchSample mysample;
        fillSample (mysample, *sample, meta);
 
        std::vector<BatchSegment> subsegments;
        splitSample (mysample, subsegments);
 
        addSample (myjob, mysample, subsegments);
      }
    }
  }
 
 
 
  void BatchDriver ::
  testInvariant () const
  {}
 
 
 
  BatchDriver ::
  BatchDriver ()
  {
    RCU_NEW_INVARIANT (this);
  }
 
 
 
  ::StatusCode BatchDriver ::
  doManagerStep (Detail::ManagerData& data) const
  {
    using namespace msgEventLoop;
    ANA_CHECK (Driver::doManagerStep (data));
    switch (data.step)
    {
    case Detail::ManagerStep::fillOptions:
      {
        data.sharedFileSystem
          = data.options.castBool (Job::optBatchSharedFileSystem, true);
      }
      break;
 
    case Detail::ManagerStep::createSubmitDir:
      {
        if (data.sharedFileSystem) // Shared file-system, write to output
          data.batchWriteLocation = data.submitDir;
        else
          data.batchWriteLocation = ".";
 
        if (data.sharedFileSystem) // Shared file-system, use local path
          data.batchSubmitLocation = data.submitDir+"/submit";
        else
          data.batchSubmitLocation = ".";
      }
      break;
 
    case Detail::ManagerStep::updateOutputLocation:
      {
        const std::string writeLocation=data.batchWriteLocation;
        for (Job::outputMIter out = data.job->outputBegin(),
               end = data.job->outputEnd(); out != end; ++ out)
        {
          if (out->output() == nullptr)
          {
            out->output (new SH::DiskOutputLocal
                         (writeLocation + "/fetch/data-" + out->label() + "/"));
          }
        }
      }
      break;
 
    case Detail::ManagerStep::batchCreateDirectories:
      {
        ANA_MSG_DEBUG ("submitting batch job in location " << data.submitDir);
        const std::string submitDir = data.submitDir + "/submit";
        if (gSystem->MakeDirectory (submitDir.c_str()) != 0)
        {
          ANA_MSG_ERROR ("failed to create directory " + submitDir);
          return ::StatusCode::FAILURE;
        }
        const std::string runDir = data.submitDir + "/run";
        if (gSystem->MakeDirectory (runDir.c_str()) != 0)
        {
          ANA_MSG_ERROR ("failed to create directory " + runDir);
          return ::StatusCode::FAILURE;
        }
        const std::string fetchDir = data.submitDir + "/fetch";
        if (gSystem->MakeDirectory (fetchDir.c_str()) != 0)
        {
          ANA_MSG_ERROR ("failed to create directory " + fetchDir);
          return ::StatusCode::FAILURE;
        }
        const std::string statusDir = data.submitDir + "/status";
        if (gSystem->MakeDirectory (statusDir.c_str()) != 0)
        {
          ANA_MSG_ERROR ("failed to create directory " + statusDir);
          return ::StatusCode::FAILURE;
        }
      }
      break;
 
    case Detail::ManagerStep::batchCreateJob:
      {
        data.batchJob = std::make_unique<BatchJob> ();
        fillFullJob (*data.batchJob, *data.job, data.submitDir, data.options);
        data.batchJob->location=data.batchWriteLocation;
        {
          std::string path = data.submitDir + "/submit/config.root";
          std::unique_ptr<TFile> file (TFile::Open (path.c_str(), "RECREATE"));
          if (file == nullptr || file->IsZombie())
          {
            ANA_MSG_ERROR ("failed to open " << path << " for writing");
            return ::StatusCode::FAILURE;
          }
          // write into this file explicitly; TObject::Write would target
          // the current gDirectory, silently dropping the config if the
          // open above had failed.
          if (file->WriteObject (data.batchJob.get(), "job") <= 0)
          {
            ANA_MSG_ERROR ("failed to write job configuration to " << path);
            return ::StatusCode::FAILURE;
          }
        }
        {
          std::ofstream file ((data.submitDir + "/submit/segments").c_str());
          for (std::size_t iter = 0, end = data.batchJob->segments.size();
               iter != end; ++ iter)
          {
            file << iter << " " << data.batchJob->segments[iter].fullName << "\n";
          }
        }
      }
      break;
 
    case Detail::ManagerStep::batchScriptVar:
      {
        data.batchName = "run";
        data.batchInit = "";
        data.batchJobId = "EL_JOBID=$1\n";
      }
      break;
 
    case Detail::ManagerStep::batchMakeScript:
      {
        makeScript (data, data.batchJob->segments.size());
      }
      break;
 
    case Detail::ManagerStep::batchMakeIndices:
      {
        for (std::size_t index = 0; index != data.batchJob->segments.size(); ++ index)
          data.batchJobIndices.push_back (index);
      }
      break;
 
    case Detail::ManagerStep::readConfigResubmit:
    case Detail::ManagerStep::readConfigRetrieve:
      {
        ANA_MSG_INFO ("retrieving batch job in location " << data.submitDir);
 
        std::unique_ptr<TFile> file
          {TFile::Open ((data.submitDir + "/submit/config.root").c_str(), "READ")};
        if (file == nullptr || file->IsZombie())
        {
          ANA_MSG_ERROR ("failed to read config.root");
          return ::StatusCode::FAILURE;
        }
        data.batchJob.reset (dynamic_cast<BatchJob*>(file->Get ("job")));
        if (data.batchJob == nullptr)
        {
          ANA_MSG_ERROR ("failed to get job object from config.root");
          return ::StatusCode::FAILURE;
        }
        data.job = &data.batchJob->job;
      }
      break;
 
    case Detail::ManagerStep::batchJobStatusResubmit:
    case Detail::ManagerStep::batchJobStatusRetrieve:
      {
        for (std::size_t job = 0; job != data.batchJob->segments.size(); ++ job)
        {
          std::ostringstream completedFile;
          completedFile << data.submitDir << "/status/completed-" << job;
          const bool hasCompleted =
            (gSystem->AccessPathName (completedFile.str().c_str()) == 0);
 
          std::ostringstream failFile;
          failFile << data.submitDir << "/status/fail-" << job;
          const bool hasFail =
            (gSystem->AccessPathName (failFile.str().c_str()) == 0);
 
          if (hasCompleted)
          {
            if (hasFail)
            {
              ANA_MSG_ERROR ("sub-job " << job << " reported both success and failure");
              return ::StatusCode::FAILURE;
            } else
              data.batchJobSuccess.insert (job);
          } else
          {
            if (hasFail)
              data.batchJobFailure.insert (job);
            else
              data.batchJobUnknown.insert (job);
          }
        }
        ANA_MSG_INFO ("current job status: " << data.batchJobSuccess.size() << " success, " << data.batchJobFailure.size() << " failure, " << data.batchJobUnknown.size() << " running/unknown");
      }
      break;
 
    case Detail::ManagerStep::batchPreResubmit:
      {
        bool all_missing = false;
        if (data.resubmitOption == "ALL_MISSING")
        {
          all_missing = true;
        } else if (!data.resubmitOption.empty())
        {
          ANA_MSG_ERROR ("unknown resubmit option " + data.resubmitOption);
          return ::StatusCode::FAILURE;
        }
 
        for (std::size_t segment = 0; segment != data.batchJob->segments.size(); ++ segment)
        {
          if (all_missing)
          {
            if (data.batchJobSuccess.find (segment) == data.batchJobSuccess.end())
              data.batchJobIndices.push_back (segment);
          } else
          {
            if (data.batchJobFailure.find (segment) != data.batchJobFailure.end())
              data.batchJobIndices.push_back (segment);
          }
        }
 
        if (data.batchJobIndices.empty())
        {
          ANA_MSG_INFO ("found no jobs to resubmit");
          data.nextStep = Detail::ManagerStep::final;
          return ::StatusCode::SUCCESS;
        }
 
        for (std::size_t segment : data.batchJobIndices)
        {
          std::ostringstream command;
          command << "rm -rf";
          command << " " << RCU::Shell::quote (data.submitDir) << "/status/completed-" << segment;
          command << " " << RCU::Shell::quote (data.submitDir) << "/status/fail-" << segment;
          command << " " << RCU::Shell::quote (data.submitDir) << "/status/done-" << segment;
          RCU::Shell::exec (command.str());
        }
        data.options = *data.batchJob->job.options();
      }
      break;
 
    case Detail::ManagerStep::doRetrieve:
      {
        bool merged = mergeHists (data);
        if (merged)
        {
          diskOutputSave (data);
        }
        data.retrieved = true;
        data.completed = merged;
      }
      break;
 
    default:
      (void) true; // safe to do nothing
    }
    return ::StatusCode::SUCCESS;
  }
 
 
 
  std::string BatchDriver ::
  defaultReleaseSetup (const Detail::ManagerData& data) const
  {
    RCU_READ_INVARIANT (this);
 
    // name of tarball being made (this needs to match CondorDriver.cxx)
    const std::string tarballName("AnalysisPackage.tar.gz");
 
    std::ostringstream file;
 
    // <path of build dir>/x86_64-slc6-gcc62-opt (comes from CMake, we need this)
    const char *WORKDIR_DIR         = getenv ("WorkDir_DIR");
    // As a backup, keep the CMAKE_PREFIX_PATH (only needed, and only read,
    // when $WorkDir_DIR is not set; std::string(nullptr) would be UB)
    std::string CMAKE_DIR_str;
    if (WORKDIR_DIR == nullptr){
      msgEventLoop::ANA_MSG_INFO ("Could not find environment variable $WorkDir_DIR");
      const char *CMAKE_PREFIX_PATH = getenv ("CMAKE_PREFIX_PATH");
      if (CMAKE_PREFIX_PATH == nullptr){
        throw std::runtime_error ("neither $WorkDir_DIR nor $CMAKE_PREFIX_PATH is set, cannot locate the release");
      }
      // Instead, build from the first path in CMAKE_PREFIX_PATH
      CMAKE_DIR_str = CMAKE_PREFIX_PATH;
      if (CMAKE_DIR_str.find(":") != std::string::npos){
        // Erase everything from the colon onwards
        CMAKE_DIR_str.erase( CMAKE_DIR_str.find(":") , std::string::npos );
      }
      // Provide the remainder of the string to the workdir
      WORKDIR_DIR = CMAKE_DIR_str.data();
    }
 
    if(!data.sharedFileSystem)
    {
      file << "mkdir -p build && tar -C build/ -xf " << tarballName << " || abortJob\n";
      file << "\n";
    }
 
 
    file << "\n";
    // /cvmfs/atlas.cern.ch/repo/ATLASLocalRootBase/x86_64/AtlasSetup/.config/.asetup.site
    if(getenv("AtlasSetupSite"))  file << "export AtlasSetupSite=" << getenv("AtlasSetupSite") << "\n";
    // /cvmfs/atlas.cern.ch/repo/ATLASLocalRootBase/x86_64/AtlasSetup/V00-07-75/AtlasSetup
    if(getenv("AtlasSetup"))      file << "export AtlasSetup=" << getenv("AtlasSetup") << "\n";
    // for now, needed because of errors like:
    //  /cvmfs/atlas.cern.ch/repo/ATLASLocalRootBase/swConfig/asetup/asetupEpilog.sh: line 38:
    //      /swConfig/python/pythonFix-Linux.sh: No such file or directory
    file << "export ATLAS_LOCAL_ROOT_BASE=/cvmfs/atlas.cern.ch/repo/ATLASLocalRootBase\n";
    file << "source $ATLAS_LOCAL_ROOT_BASE/user/atlasLocalSetup.sh --quiet\n";
 
    // default setup command
    std::ostringstream defaultSetupCommand;
    {
      // AnalysisBase
      if(getenv("AtlasProject"))     defaultSetupCommand << "export AtlasProject=" << getenv("AtlasProject") << "\n";
      // 21.2.3
      if(getenv("AtlasVersion"))     defaultSetupCommand << "export AtlasVersion=" << getenv("AtlasVersion") << "\n";
      // 2017-08-16T2249 (only set if using a nightly release)
      if(getenv("AtlasBuildStamp"))  defaultSetupCommand << "export AtlasBuildStamp=" << getenv("AtlasBuildStamp") << "\n";
      // 21.2
      if(getenv("AtlasBuildBranch")) defaultSetupCommand << "export AtlasBuildBranch=" << getenv("AtlasBuildBranch") << "\n";
      // stable vs nightly
      if(getenv("AtlasReleaseType")) defaultSetupCommand << "export AtlasReleaseType=" << getenv("AtlasReleaseType") << "\n";
      defaultSetupCommand << "if [ \"${AtlasReleaseType}\" == \"stable\" ]; then\n";
      defaultSetupCommand << "     source ${AtlasSetup}/scripts/asetup.sh ${AtlasProject},${AtlasVersion} || abortJob\n";
      defaultSetupCommand << "else\n";
      defaultSetupCommand << "     source ${AtlasSetup}/scripts/asetup.sh ${AtlasProject},${AtlasBuildBranch},${AtlasBuildStamp} || abortJob\n";
      defaultSetupCommand << "fi\n";
      defaultSetupCommand << "echo \"Using default setup command\"";
    }
 
    file << data.options.castString(Job::optBatchSetupCommand, defaultSetupCommand.str()) << " || abortJob\n";
    if(data.sharedFileSystem) file << "source " << WORKDIR_DIR << "/setup.sh || abortJob\n";
    else                 file << "source build/setup.sh || abortJob\n";
    file << "\n";
 
    if(!data.sharedFileSystem)
    {
      std::ostringstream cmd;
      //cmd << "tar --dereference -C " << WORKDIR_DIR << " -czf " << tarballName << " .";
      cmd << "cpack -D CPACK_INSTALL_PREFIX=. -G TGZ --config $TestArea/CPackConfig.cmake";
 
      // suppress the output from the command
      if (gSystem->Exec (cmd.str().c_str()) != 0){
        throw std::runtime_error ("failed to execute: " + cmd.str());
      }
 
      const char *WorkDir_VERSION = std::getenv ("WorkDir_VERSION");
      const char *WorkDir_PLATFORM = std::getenv ("WorkDir_PLATFORM");
      if (WorkDir_VERSION == nullptr || WorkDir_PLATFORM == nullptr){
        throw std::runtime_error ("$WorkDir_VERSION and $WorkDir_PLATFORM must both be set to build the release tarball");
      }
      std::ostringstream mv_command;
      mv_command << "mv WorkDir_" << WorkDir_VERSION << "_" << WorkDir_PLATFORM << ".tar.gz " << tarballName;
      if (gSystem->Exec (mv_command.str().c_str()) != 0){
        throw std::runtime_error ("failed to execute: " + mv_command.str());
      }
    }
 
    return file.str();
  }
 
 
 
  void BatchDriver ::
  makeScript (Detail::ManagerData& data,
              std::size_t njobs) const
  {
    RCU_READ_INVARIANT (this);
 
    std::string name = data.batchName;
    bool multiFile = (name.find ("{JOBID}") != std::string::npos);
 
    // Build the release setup once, outside the per-script loop.  The text
    // is the same for every script, and defaultReleaseSetup also builds the
    // release tarball as a side effect (for a non-shared file system), which
    // must not be repeated once per generated script.
    const std::string releaseSetup =
      data.batchSkipReleaseSetup ? std::string() : defaultReleaseSetup (data);
 
    for (std::size_t index = 0, end = multiFile ? njobs : 1; index != end; ++ index)
    {
      std::ostringstream str;
      str << index;
      const std::string fileName = data.submitDir + "/submit/" + RCU::substitute (name, "{JOBID}", str.str());
 
      {
        std::ofstream file (fileName.c_str());
        file << "#!/bin/bash\n";
        file << "echo starting batch job initialization\n";
        file << RCU::substitute (data.batchInit, "{JOBID}", str.str()) << "\n";
        file << "echo batch job user initialization finished\n";
        if (multiFile) file << "EL_JOBID=" << index << "\n\n";
        else           file << data.batchJobId << "\n";
 
        file << "function abortJob {\n";
        file << "  echo \"abort EL_JOBID=${EL_JOBID}\"\n";
        file << "  touch \"" << data.batchWriteLocation << "/status/fail-$EL_JOBID\"\n";
        file << "  touch \"" << data.batchWriteLocation << "/status/done-$EL_JOBID\"\n";
        file << "  exit 1\n";
        file << "}\n\n";
 
 
        file << "EL_JOBSEG=`grep \"^$EL_JOBID \" \"" << data.batchSubmitLocation << "/segments\" | awk ' { print $2 }'`\n";
        file << "test \"$EL_JOBSEG\" != \"\" || abortJob\n";
        file << "hostname\n";
        file << "pwd\n";
        file << "whoami\n";
        file << shellInit << "\n";
 
        if(!data.sharedFileSystem)
        { // Create output transfer directories
          file << "mkdir \"fetch\" || abortJob\n";
          file << "mkdir \"status\" || abortJob\n";
          file << "\n";
        }
 
        if(data.sharedFileSystem)
        {
          file << "test \"$TMPDIR\" == \"\" && TMPDIR=/tmp\n";
          file << "RUNDIR=${TMPDIR}/EventLoop-Worker-$EL_JOBSEG-`date +%s`-$$\n";
          file << "mkdir \"$RUNDIR\" || abortJob\n";
          file << "cd \"$RUNDIR\" || abortJob\n";
        }
 
        if (!data.batchSkipReleaseSetup)
          file << releaseSetup;
 
        file << "eventloop_batch_worker $EL_JOBID '" << data.batchSubmitLocation << "/config.root' || abortJob\n";
 
        file << "test -f \"" << data.batchWriteLocation << "/status/completed-$EL_JOBID\" || "
             << "touch \"" << data.batchWriteLocation << "/status/fail-$EL_JOBID\"\n";
        file << "touch \"" << data.batchWriteLocation << "/status/done-$EL_JOBID\"\n";
        if(data.sharedFileSystem) file << "cd .. && rm -rf \"$RUNDIR\"\n";
      }
 
      {
        std::ostringstream cmd;
        cmd << "chmod +x " << RCU::Shell::quote (fileName);
        if (gSystem->Exec (cmd.str().c_str()) != 0)
          throw std::runtime_error ("failed to execute: " + cmd.str());
      }
    }
  }
 
 
 
  bool BatchDriver ::
  mergeHists (Detail::ManagerData& data)
  {
    using namespace msgEventLoop;
 
    // This picks up the DiskOutput object used to write out our
    // histograms, we will then use that to locate the output files.
    // this is not really the best way of doing this, but there are
    // bigger rewrites of this code excepted, so I don't want to spend
    // a lot of time on this now (06 Feb 19).
    std::unique_ptr<SH::DiskOutput> origHistOutputMemory;
    const SH::DiskOutput *origHistOutput = nullptr;
    for (auto iter = data.job->outputBegin(), end = data.job->outputEnd();
         iter != end; ++ iter)
    {
      if (iter->label() == Job::histogramStreamName)
        origHistOutput = iter->output();
    }
    if (origHistOutput == nullptr)
    {
      origHistOutputMemory = std::make_unique<SH::DiskOutputLocal>
        (data.submitDir + "/fetch/hist-");
      origHistOutput = origHistOutputMemory.get();
    }
    RCU_ASSERT (origHistOutput != nullptr);
 
    bool result = true;
 
    ANA_MSG_DEBUG ("merging histograms in location " << data.submitDir);
 
    for (std::size_t sample = 0, end = data.batchJob->samples.size();
         sample != end; ++ sample)
    {
      const BatchSample& mysample (data.batchJob->samples[sample]);
 
      std::ostringstream output;
      output << data.submitDir << "/hist-" << data.batchJob->samples[sample].name << ".root";
      if (gSystem->AccessPathName (output.str().c_str()) != 0)
      {
        ANA_MSG_VERBOSE ("merge files for sample " << data.batchJob->samples[sample].name);
 
        bool complete = true;
        std::vector<std::string> input;
        for (std::size_t segment = mysample.begin_segments,
               end = mysample.end_segments; segment != end; ++ segment)
        {
          const BatchSegment& mysegment = data.batchJob->segments[segment];
 
          const std::string hist_file = origHistOutput->targetURL
            (mysegment.sampleName, mysegment.segmentName, ".root");
 
          ANA_MSG_VERBOSE ("merge segment " << segment << " completed=" << (data.batchJobSuccess.find(segment)!=data.batchJobSuccess.end()) << " fail=" << (data.batchJobFailure.find(segment)!=data.batchJobFailure.end()) << " unknown=" << (data.batchJobUnknown.find(segment)!=data.batchJobUnknown.end()));
 
          input.push_back (hist_file);
 
          if (data.batchJobFailure.find(segment)!=data.batchJobFailure.end())
          {
            std::ostringstream message;
            message << "subjob " << segment << "/" << mysegment.fullName
                    << " failed";
            throw std::runtime_error (message.str());
          }
          else if (data.batchJobSuccess.find(segment)==data.batchJobSuccess.end())
            complete = false, result = false;
        }
        if (complete)
        {
          RCU::hadd (output.str(), input);
 
          // Merge output data directories
          for (Job::outputIter out = data.batchJob->job.outputBegin(),
                 end = data.batchJob->job.outputEnd(); out != end; ++ out)
            {
              output.str("");
              output << data.submitDir << "/data-" << out->label();
 
              if(gSystem->AccessPathName(output.str().c_str()))
                gSystem->mkdir(output.str().c_str(),true);
 
 
              output << "/" << data.batchJob->samples[sample].name << ".root";
 
              std::vector<std::string> dataInput;
              for (std::size_t segment = mysample.begin_segments,
                     dataEnd = mysample.end_segments; segment != dataEnd; ++ segment)
                {
                  const BatchSegment& mysegment = data.batchJob->segments[segment];
 
                  const std::string infile =
                    data.submitDir + "/fetch/data-" + out->label() + "/" + mysegment.fullName + ".root";
 
                  dataInput.push_back (infile);
                }
 
              RCU::hadd(output.str(), dataInput);
            }
        }
      }
    }
    return result;
  }
}