RNTupleParallelWriter.hxx

Go to the documentation of this file.

/// \file ROOT/RNTupleParallelWriter.hxx
/// \author Jonas Hahnfeld <jonas.hahnfeld@cern.ch>
/// \date 2024-02-01
 
/*************************************************************************
 * Copyright (C) 1995-2024, Rene Brun and Fons Rademakers.               *
 * All rights reserved.                                                  *
 *                                                                       *
 * For the licensing terms see $ROOTSYS/LICENSE.                         *
 * For the list of contributors see $ROOTSYS/README/CREDITS.             *
 *************************************************************************/
 
#ifndef ROOT_RNTupleParallelWriter
#define ROOT_RNTupleParallelWriter
 
#include <ROOT/RNTupleMetrics.hxx>
#include <ROOT/RNTupleWriteOptions.hxx>
 
#include <memory>
#include <mutex>
#include <string_view>
#include <vector>
 
class TDirectory;
 
namespace ROOT {
 
class RNTupleModel;
 
namespace Internal {
class RPageSink;
} // namespace Internal
 
class RNTupleFillContext;
 
/**
\class ROOT::RNTupleParallelWriter
\ingroup NTuple
\brief A writer to fill an RNTuple from multiple contexts
 
Compared to the sequential RNTupleWriter, a parallel writer enables the creation of multiple RNTupleFillContext (see
CreateFillContext()).  Each fill context prepares independent clusters that are appended to the common RNTuple with
internal synchronization.  Before destruction, all fill contexts must have flushed their data and been destroyed (or
data could be lost!).
 
For user convenience, CreateFillContext() is thread-safe and may be called from multiple threads in parallel at any
time, also after some data has already been written.  Internally, the original model is cloned and ownership is passed
to a newly created RNTupleFillContext.  For that reason, it is recommended to use RNTupleModel::CreateBare when creating
the model for parallel writing and avoid the allocation of a useless default REntry per context.
 
Note that the sequence of independently prepared clusters is indeterminate and therefore entries are only partially
ordered:  Entries from one context are totally ordered as they were filled.  However, there is no orderering with other
contexts and the entries may be appended to the RNTuple either before or after other entries written in parallel into
other contexts.  In addition, two consecutive entries in one fill context can end up separated in the final RNTuple, if
they happen to fall onto a cluster boundary and other contexts append more entries before the next cluster is full.
 
At the moment, the parallel writer does not (yet) support incremental updates of the underlying model. Please refer to
RNTupleWriter::CreateModelUpdater if required for your use case.
*/
class RNTupleParallelWriter {
private:
   /// A global mutex to protect the internal data structures of this object.
   std::mutex fMutex;
   /// A mutex to synchronize the final page sink.
   std::mutex fSinkMutex;
   /// The final RPageSink that represents the synchronization point.
   std::unique_ptr<ROOT::Internal::RPageSink> fSink;
   /// The original RNTupleModel connected to fSink; needs to be destructed before it.
   std::unique_ptr<ROOT::RNTupleModel> fModel;
   Experimental::Detail::RNTupleMetrics fMetrics;
   /// List of all created helpers. They must be destroyed before this RNTupleParallelWriter is destructed.
   std::vector<std::weak_ptr<RNTupleFillContext>> fFillContexts;
 
   RNTupleParallelWriter(std::unique_ptr<ROOT::RNTupleModel> model, std::unique_ptr<ROOT::Internal::RPageSink> sink);
   RNTupleParallelWriter(const RNTupleParallelWriter &) = delete;
   RNTupleParallelWriter &operator=(const RNTupleParallelWriter &) = delete;
   RNTupleParallelWriter(RNTupleParallelWriter &&) = delete;
   RNTupleParallelWriter &operator=(RNTupleParallelWriter &&) = delete;
 
public:
   /// Recreate a new file and return a writer to write an RNTuple.
   static std::unique_ptr<RNTupleParallelWriter>
   Recreate(std::unique_ptr<ROOT::RNTupleModel> model, std::string_view ntupleName, std::string_view storage,
            const ROOT::RNTupleWriteOptions &options = ROOT::RNTupleWriteOptions());
   /// Append an RNTuple to the existing file.
   ///
   /// While the writer synchronizes between multiple fill contexts created from the same writer, there is no
   /// synchronization with other writers or other clients that write into the same file. The caller must ensure that
   /// the underlying file is not be accessed while data is filled into any created context. To improve performance, it
   /// is allowed to use special methods that are guaranteed to not interact with the underlying file, such as
   /// RNTupleFillContext::FillNoFlush().
   static std::unique_ptr<RNTupleParallelWriter>
   Append(std::unique_ptr<ROOT::RNTupleModel> model, std::string_view ntupleName, TDirectory &fileOrDirectory,
          const ROOT::RNTupleWriteOptions &options = ROOT::RNTupleWriteOptions());
 
   ~RNTupleParallelWriter();
 
   /// Create a new RNTupleFillContext that can be used to fill entries and prepare clusters in parallel. This method is
   /// thread-safe and may be called from multiple threads in parallel at any time, also after some data has already
   /// been written.
   ///
   /// Note that all fill contexts must be destroyed before CommitDataset() is called.
   std::shared_ptr<RNTupleFillContext> CreateFillContext();
 
   /// Automatically called by the destructor
   void CommitDataset();
 
   void EnableMetrics() { fMetrics.Enable(); }
   const Experimental::Detail::RNTupleMetrics &GetMetrics() const { return fMetrics; }
};
 
} // namespace ROOT
 
#endif