SerializerH5.hpp File Reference

MPI-parallel HDF5 serializer implementing the SerializerBase interface. More...

#include "SerializerBase.hpp"
#include <hdf5.h>
#include <fstream>
#include <map>
#include <set>

Include dependency graph for SerializerH5.hpp:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
class	DNDS::Serializer::SerializerH5
	MPI-parallel HDF5 serializer; all ranks collectively read/write a single .h5 file. More...

Namespaces
namespace	DNDS
	the host side operators are provided as implemented
namespace	DNDS::Serializer

Detailed Description

MPI-parallel HDF5 serializer implementing the SerializerBase interface.

Unit Test Coverage (test_Serializer.cpp, MPI np=1,2,4)

• Scalar round-trip: WriteInt/ReadInt, WriteIndex/ReadIndex, WriteReal/ReadReal, WriteString/ReadString
• Vector round-trip: WriteRealVector, WriteIndexVector with explicit ArrayGlobalOffset; ReadRealVector/ReadIndexVector with ArrayGlobalOffset_Unknown
• Distributed vector: non-uniform per-rank sizes, write with explicit offset, read with both ArrayGlobalOffset_Unknown (auto-detect from ::rank_offsets) and explicit offset
• uint8 distributed round-trip: two-pass read (nullptr size query, then read)
• Path operations: CreatePath, GoToPath, GetCurrentPath, ListCurrentPath (groups materialized by writing content), WriteInt/ReadInt on nested paths
• String round-trip: WriteString/ReadString with fixed-length HDF5 attributes

Collective I/O and zero-size partitions

All Read/Write vector and byte-array methods use MPI-collective HDF5 calls. Every rank must call them in the same order, even when its local element count is 0 (which happens when nGlobal < nRanks under EvenSplit).

Internally, ReadDataVector uses a two-pass pattern:

• Pass 1 (buf == nullptr): queries dataset size and resolves the offset.
• Pass 2 (buf != nullptr): performs the collective H5Dread.

When local size is 0, std::vector<>::data() / host_device_vector<>::data() may return nullptr, which would skip the H5Dread block (guarded by if (buf != nullptr)) and hang the other ranks. To prevent this, each Read*Vector / ReadShared*Vector caller passes a dummy stack pointer when size == 0. Callers of ReadUint8Array must do the same (see SerializerBase).

H5_ReadDataset and H5_WriteDataset accept nLocal == 0: the hyperslab selection with count == 0 selects nothing, so no data is transferred, but the rank still participates in the collective call.

Not Yet Tested

• SetChunkAndDeflate impact, SetCollectiveRW impact
• WriteSharedIndexVector / ReadSharedIndexVector (H5 deduplication)
• WriteRowsizeVector / ReadRowsizeVector
• WriteIndexVectorPerRank
• ArrayGlobalOffset_One semantics (rank-0-only write)
• ArrayGlobalOffset_Parts auto-offset

Definition in file SerializerH5.hpp.