RNTupleWriteOptions.hxx

Go to the documentation of this file.

/// \file ROOT/RNTupleWriteOptions.hxx
/// \ingroup NTuple
/// \author Jakob Blomer <jblomer@cern.ch>
/// \date 2024-02-22
 
/*************************************************************************
 * Copyright (C) 1995-2019, 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_RNTupleWriteOptions
#define ROOT_RNTupleWriteOptions
 
#include <Compression.h>
 
#include <cstdint>
#include <cstddef>
#include <memory>
 
namespace ROOT {
 
class RNTupleWriteOptions;
 
namespace Internal {
 
class RNTupleWriteOptionsManip final {
public:
   static void SetMaxKeySize(RNTupleWriteOptions &options, std::uint64_t maxKeySize);
};
 
} // namespace Internal
 
// clang-format off
/**
\class ROOT::RNTupleWriteOptions
\ingroup NTuple
\brief Common user-tunable settings for storing RNTuples
 
All page sink classes need to support the common options.
 
<table>
<tr>
<th>Option name</th>
<th>Type</th>
<th>Default</th>
<th>Description</th>
</tr>
 
<tr>
<td>`Compression`</td>
<td>`std::uint32_t`</td>
<td>RCompressionSetting::EDefaults::kUseGeneralPurpose</td>
<td>
The compression settings for this RNTuple
</td>
</tr>
 
<tr>
<td>`ApproxZippedClusterSize`</td>
<td>`std::size_t`</td>
<td>128 MiB</td>
<td>
Approximation of the target compressed cluster size
</td>
</tr>
 
<tr>
<td>`MaxUnzippedClusterSize`</td>
<td>`std::size_t`</td>
<td>1280 MiB</td>
<td>
Memory limit for committing a cluster: with very high compression ratio, we need a limit
on how large the I/O buffer can grow during writing.
</td>
</tr>
 
<tr>
<td>`InitialUnzippedPageSize`</td>
<td>`std::size_t`</td>
<td>256</td>
<td>
Initially, columns start with a page of this size. The default value is chosen to accomodate at least 32 elements
of 64 bits, or 64 elements of 32 bits. If more elements are needed, pages are increased up until the byte limit
given by the option `MaxUnzippedPageSize` or until the total page buffer limit is reached (as a sum of all page buffers).
The total write buffer limit needs to be large enough to hold the initial pages of all columns.
</td>
</tr>
 
<tr>
<td>`MaxUnzippedPageSize`</td>
<td>`std::size_t`</td>
<td>1 MiB</td>
<td>
Pages can grow only to the given limit in bytes.
</td>
</tr>
 
<tr>
<td>`PageBufferBudget`</td>
<td>`std::size_t`</td>
<td>0 / auto</td>
<td>
The maximum size that the sum of all page buffers used for writing into a persistent sink are allowed to use.
If set to zero, RNTuple will auto-adjust the budget based on the value of `ApproxZippedClusterSize`.
If set manually, the size needs to be large enough to hold all initial page buffers.
The total amount of memory for writing is larger, e.g. for the additional compressed buffers etc.
Use RNTupleModel::EstimateWriteMemoryUsage() for the total estimated memory use for writing.
The default values are tuned for a total write memory of around 400 MiB per fill context.
</td>
</tr>
 
<tr>
<td>`UseBufferedWrite`</td>
<td>`bool`</td>
<td>`true`</td>
<td>
Whether to use buffered writing (with RPageSinkBuf). This buffers compressed pages in memory, reorders them
to keep pages of the same column adjacent, and coalesces the writes when committing a cluster.
</td>
</tr>
 
<tr>
<td>`UseDirectIO`</td>
<td>`bool`</td>
<td>`false`</td>
<td>
Whether to use Direct I/O for writing. Note that this introduces alignment requirements that may very between
filesystems and platforms.
</td>
</tr>
 
<tr>
<td>`WriteBufferSize`</td>
<td>`std::size_t`</td>
<td>4 MiB</td>
<td>
Buffer size to use for writing to files, must be a multiple of 4096 bytes. Testing suggests that 4MiB gives best
performance (with Direct I/O) at a reasonable memory consumption.
</td>
</tr>
 
<tr>
<td>`UseImplicitMT`</td>
<td>EImplicitMT</td>
<td>EImplicitMT::kDefault</td>
<td>
Whether to use implicit multi-threading to compress pages. Only has an effect if buffered writing is turned on.
The meaning of EImplicitMT::kDefault depends on the used writer: For the (sequential) RNTupleWriter, it translates
to EImplicitMT::kOn and the user has to manually disable the use of implicit multi-threading if it is not wanted.
For the RNTupleParalellWriter, the implementation defaults to EImplicitMT::kOff in order to avoid interference with
explicit parallelism that might create one RNTupleFillContext per thread. If implicit multi-threading is wanted on
top of this, the user has to explicitly request EImplicitMT::kOn.
</td>
</tr>
 
<tr>
<td>`EnablePageChecksums`</td>
<td>`bool`</td>
<td>`true`</td>
<td>
If set, checksums will be calculated and written for every page.
If turned off, will also turn off `EnableSamePageMerging`.
</td>
</tr>
 
<tr>
<td>`EnableSamePageMerging`</td>
<td>`bool`</td>
<td>`true`</td>
<td>
If set, identical pages are deduplicated and aliased on disk.
Requires `EnablePageChecksums` and will throw if previously disabled.
</td>
</tr>
 
</table>
*/
// clang-format on
class RNTupleWriteOptions {
public:
   enum class EImplicitMT {
      kOff,
      kOn,
      kDefault,
   };
 
   // clang-format off
   static constexpr std::uint64_t kDefaultMaxKeySize = 0x4000'0000; // 1 GiB
   // clang-format on
 
   friend Internal::RNTupleWriteOptionsManip;
 
protected:
   std::uint32_t fCompression{RCompressionSetting::EDefaults::kUseGeneralPurpose};
   std::size_t fApproxZippedClusterSize = 128 * 1024 * 1024;
   std::size_t fMaxUnzippedClusterSize = 10 * fApproxZippedClusterSize;
   std::size_t fInitialUnzippedPageSize = 256;
   std::size_t fMaxUnzippedPageSize = 1024 * 1024;
   std::size_t fPageBufferBudget = 0;
   bool fUseBufferedWrite = true;
   bool fUseDirectIO = false;
   std::size_t fWriteBufferSize = 4 * 1024 * 1024;
   EImplicitMT fUseImplicitMT = EImplicitMT::kDefault;
   bool fEnablePageChecksums = true;
   bool fEnableSamePageMerging = true;
   /// Specifies the max size of a payload storeable into a single TKey. When writing an RNTuple to a ROOT file,
   /// any payload whose size exceeds this will be split into multiple keys.
   std::uint64_t fMaxKeySize = kDefaultMaxKeySize;
 
public:
 
   virtual ~RNTupleWriteOptions() = default;
   virtual std::unique_ptr<RNTupleWriteOptions> Clone() const;
 
   std::uint32_t GetCompression() const { return fCompression; }
   void SetCompression(std::uint32_t val) { fCompression = val; }
   void SetCompression(RCompressionSetting::EAlgorithm::EValues algorithm, int compressionLevel)
   {
      fCompression = CompressionSettings(algorithm, compressionLevel);
   }
 
   std::size_t GetApproxZippedClusterSize() const { return fApproxZippedClusterSize; }
   void SetApproxZippedClusterSize(std::size_t val);
 
   std::size_t GetMaxUnzippedClusterSize() const { return fMaxUnzippedClusterSize; }
   void SetMaxUnzippedClusterSize(std::size_t val);
 
   std::size_t GetInitialUnzippedPageSize() const { return fInitialUnzippedPageSize; }
   void SetInitialUnzippedPageSize(std::size_t val);
 
   std::size_t GetMaxUnzippedPageSize() const { return fMaxUnzippedPageSize; }
   void SetMaxUnzippedPageSize(std::size_t val);
 
   std::size_t GetPageBufferBudget() const;
   void SetPageBufferBudget(std::size_t val) { fPageBufferBudget = val; }
 
   bool GetUseBufferedWrite() const { return fUseBufferedWrite; }
   void SetUseBufferedWrite(bool val) { fUseBufferedWrite = val; }
 
   bool GetUseDirectIO() const { return fUseDirectIO; }
   void SetUseDirectIO(bool val) { fUseDirectIO = val; }
 
   std::size_t GetWriteBufferSize() const { return fWriteBufferSize; }
   void SetWriteBufferSize(std::size_t val) { fWriteBufferSize = val; }
 
   EImplicitMT GetUseImplicitMT() const { return fUseImplicitMT; }
   void SetUseImplicitMT(EImplicitMT val) { fUseImplicitMT = val; }
 
   bool GetEnablePageChecksums() const { return fEnablePageChecksums; }
   /// Note that turning off page checksums will also turn off the same page merging optimization (see tuning.md)
   void SetEnablePageChecksums(bool val)
   {
      fEnablePageChecksums = val;
      if (!fEnablePageChecksums) {
         fEnableSamePageMerging = false;
      }
   }
 
   bool GetEnableSamePageMerging() const { return fEnableSamePageMerging; }
   void SetEnableSamePageMerging(bool val);
 
   std::uint64_t GetMaxKeySize() const { return fMaxKeySize; }
 
   friend bool operator==(const RNTupleWriteOptions &lhs, const RNTupleWriteOptions &rhs)
   {
      return lhs.fCompression == rhs.fCompression && lhs.fApproxZippedClusterSize == rhs.fApproxZippedClusterSize &&
             lhs.fMaxUnzippedClusterSize == rhs.fMaxUnzippedClusterSize &&
             lhs.fInitialUnzippedPageSize == rhs.fInitialUnzippedPageSize &&
             lhs.fMaxUnzippedPageSize == rhs.fMaxUnzippedPageSize && lhs.fPageBufferBudget == rhs.fPageBufferBudget &&
             lhs.fUseBufferedWrite == rhs.fUseBufferedWrite && lhs.fUseDirectIO == rhs.fUseDirectIO &&
             lhs.fWriteBufferSize == rhs.fWriteBufferSize && lhs.fUseImplicitMT == rhs.fUseImplicitMT &&
             lhs.fEnablePageChecksums == rhs.fEnablePageChecksums &&
             lhs.fEnableSamePageMerging == rhs.fEnableSamePageMerging && lhs.fMaxKeySize == rhs.fMaxKeySize;
   }
 
   friend bool operator!=(const RNTupleWriteOptions &lhs, const RNTupleWriteOptions &rhs) { return !(lhs == rhs); }
};
 
namespace Internal {
inline void RNTupleWriteOptionsManip::SetMaxKeySize(RNTupleWriteOptions &options, std::uint64_t maxKeySize)
{
   options.fMaxKeySize = maxKeySize;
}
 
} // namespace Internal
} // namespace ROOT
 
#endif // ROOT_RNTupleWriteOptions