Blosc2 in BP5: invalid blosc buffer version

[franzpoeschel]

Describe the bug
Compressing with Blosc2 in BP5 will sometimes lead to unreadable datasets.

To Reproduce
I observed the bug when trying to post-hoc-compress a PIConGPU dataset with BP5. The following code snippet reads the metadata of a BP5 dataset written by PIConGPU (uncompressed.zip, metadata only, I removed the data.* files), and recreates all variables that it finds, filling them with zeroes and compression with BLOSC_BITSHUFFLE.

#include <adios2.h>
#include <adios2/common/ADIOSMacros.h>
#include <cstdlib>
#include <iostream>
#include <sstream>
#include <stdexcept>
#include <string>
#include <vector>

struct ADIOS2Params
{
    adios2::ADIOS adios;
    adios2::IO IO;
    adios2::Engine engine;
};

struct Params
{
    ADIOS2Params adios_in;
    ADIOS2Params adios_out;
    std::vector<std::unique_ptr<char[]>> buffers;
};

template <typename T>
void handle_var(Params &params, std::string const &name)
{
    auto var = params.adios_in.IO.InquireVariable<T>(name);
    if (!var)
    {
        throw std::runtime_error("Failed inquiring variable '" + name + "'");
    }
    adios2::Dims offset(var.Shape().size(), 0);
    size_t flat_extent = 1;
    for (auto ext : var.Shape())
    {
        flat_extent *= ext;
    }
    auto out_var = params.adios_out.IO.DefineVariable<T>(
        name, var.Shape(), offset, var.Shape());
    out_var.AddOperation("blosc", {{"doshuffle", "BLOSC_BITSHUFFLE"}});
    std::unique_ptr<char[]> writeBuffer{new char[flat_extent * sizeof(T)]{}};
    params.adios_out.engine.Put(
        out_var, reinterpret_cast<T *>(writeBuffer.get()));
    params.buffers.push_back(std::move(writeBuffer));
}

int main(int argsc, char **argsv)
{
    if (argsc != 3)
    {
        std::cout << "Usage: " << argsv[0]
                  << " <sample data location> <out file>" << std::endl;
        std::exit(1);
    }

    Params params;

    char const *in_file = argsv[1];
    char const *out_file = argsv[2];

    params.adios_in.adios = adios2::ADIOS{};
    params.adios_in.IO = params.adios_in.adios.DeclareIO("IO");
    params.adios_in.IO.SetEngine("bp5");
    params.adios_in.engine =
        params.adios_in.IO.Open(in_file, adios2::Mode::Read);

    params.adios_out.adios = adios2::ADIOS{};
    params.adios_out.IO = params.adios_out.adios.DeclareIO("IO");
    params.adios_out.IO.SetEngine("bp5");
    params.adios_out.engine =
        params.adios_out.IO.Open(out_file, adios2::Mode::Write);

    std::vector<char> mockHostnameTable;

    while (params.adios_in.engine.BeginStep() == adios2::StepStatus::OK)
    {
        params.adios_out.engine.BeginStep();

        for (auto const &[var, var_meta] :
             params.adios_in.IO.AvailableVariables())
        {
            std::string actualType = var_meta.at("Type");

#define SWITCH_TYPE(ctype, repr)                                               \
    if (#repr == actualType || std::string(#repr) + "_t" == actualType)        \
    {                                                                          \
        handle_var<ctype>(params, var);                                        \
    }                                                                          \
    else

            ADIOS2_FOREACH_PRIMITVE_STDTYPE_2ARGS(SWITCH_TYPE)
            {
                throw std::runtime_error("Unknown type: '" + actualType + "'");
            }
#undef SWITCH_TYPE
        }
        params.adios_in.engine.EndStep();
        params.adios_out.engine.EndStep();
        params.buffers.clear();
    }

    params.adios_in.engine.Close();
    params.adios_out.engine.Close();
}

cmake_minimum_required(VERSION 3.12.0)

project(adios2_mock_stream)

find_package(ADIOS2 REQUIRED)

foreach(binary bp5_blosc)
    add_executable(${binary} ${binary}.cpp)
    set_property(TARGET ${binary} PROPERTY CXX_STANDARD 17)

    target_link_libraries(${binary} PRIVATE adios2::cxx11)
endforeach(binary adios2_mock_stream adios2_read_stream)

Trying to read some of the variables will now fail:

$ ./bp5_blosc uncompressed_000500.bp5/ compressed.bp5
$ bpls -d compressed.bp5/ -e /data/500/particles/i/momentum/x
  float     /data/500/particles/i/momentum/x                          {164659200}

bpls caught an exception
[Fri Feb 24 14:12:10 2023] [ADIOS2 EXCEPTION] <Operator> <CompressBlosc> <InverseOperate> : invalid blosc buffer version

Expected behavior
No reading failure

Desktop (please complete the following information):

• OS/Platform: Ubuntu 22.04.1 LTS Singularity container on an Ubuntu 20.04.5 LTS system
• Build: v2.9.0-rc1, gcc 11.3.0

Additional context
The same workflow works fine with BP4.

Following up
Was the issue fixed? Please report back.

[pnorbert]

@lizdulac Any update on this? Franz still has this issue.

[vicentebolea]

Any update on this?

[ax3l]

@lizdulac @pnorbert did you have time to look into this BP5 + Blosc2 issue? :)

[franzpoeschel]

Tested today: The issue persists with the recent master branch cdd58fb

[anagainaru]

We were able to reproduce the problem but we cannot find what is causing this. It is not a straight forward issue in BP5, it might be some property of this dataset. I am still working on it.