How it works

The following image illustrates the core idea behind matrix transposition in Blosc2:

On the left, we have matrix A, and on the right, its transpose B = Aᵗ. Transposition means that an element located at position A[i][j] is moved to position B[j][i].

Each element in matrix A is shown as a colored cell, visually, the 2×3 blue chunk in A gets rotated and repositioned in B, illustrating how entire blocks of data shift when the matrix is transposed.

The red border marks the original chunk in both matrices. What’s interesting is that the shape of the chunks themselves changes during transposition. While A may use a 2×3 chunk layout, the transposed matrix B ends up with a different chunk geometry due to the new memory layout.

Why does this matter? In Blosc2, chunks define how data is split and compressed. Choosing the right chunk shape can drastically influence compression efficiency and decompression speed. So, understanding how chunks shift during operations like transposition is essential for optimizing performance.

Performance benchmark: Transposing matrices with Blosc2 vs NumPy

To evaluate the performance of the new matrix transposition implementation in Blosc2, I conducted a series of benchmarks comparing it to NumPy, which serves as the baseline due to its widespread use and high optimization level. The goal was to observe how both approaches perform when handling matrices of increasing size and to understand the impact of different chunk configurations in Blosc2.

Results and discussion

The chart below summarizes the benchmark results for matrix transposition using NumPy and Blosc2, across various chunk shapes and matrix sizes.

While NumPy sets a strong performance baseline, the behaviour of Blosc2 becomes particularly interesting when we dive into how different chunk configurations affect transposition speed. The following observations highlight how crucial the choice of chunk shape is to achieving optimal performance.

• Large, square chunks such as (1000, 1000) and (5000, 5000) offered the fastest and most stable transpositions, keeping execution times below 3 seconds even for matrices exceeding 2 GB. These shapes minimize chunk overhead and maximize cache locality, which is key to achieving high throughput in Blosc2.

• Automatic chunking also performs reasonably well, though its timings fluctuate more as matrix sizes increase. It appears that while the automatic mode is generally robust, it doesn't always choose optimal chunk boundaries, especially for very large arrays, which can lead to inconsistent transposition speeds.

• Small rectangular chunks show the poorest performance, with transposition times skyrocketing for large matrices. For example, (150, 300) reaches nearly 30 seconds at around 2200 MB of matrix size, about 100× slower than NumPy. This is likely due to poor cache utilization and increased overhead in managing a high number of small blocks.

These benchmarks confirm that chunk configuration plays a critical role in how well Blosc2 performs. Unlike traditional in-memory libraries, Blosc2’s behaviour can vary greatly depending on how data is internally structured.

Conclusion

The benchmarks highlight one key insight: Blosc2 is highly sensitive to chunk shape, and its performance can range from excellent to poor depending on how it is configured. With the right chunk size, Blosc2 can offer both high-speed transpositions and advanced features like compression and out-of-core processing. However, misconfigured chunks, especially those that are too small or uneven, can drastically reduce its effectiveness. This makes chunk tuning an essential step for anyone seeking to get the most out of Blosc2 for large-scale matrix operations.

Appendix: Unexpected behaviour in NumPy

While running the benchmarks, two unusual spikes were consistently observed in the performance of NumPy around matrices of approximately 500 MB, 1100 MB and 1900 MB in size. This can be clearly seen in the plot below:

This sudden increase in transposition time is consistently reproducible and does not seem to correlate with the gradual increase expected from larger memory sizes. We have also observed this behaviour in other machines, although at different sizes.

This observation reinforces the importance of testing under realistic and varied conditions, as performance is not always linear or intuitive.

Matrix Multiplication

Matrix multiplication is a fundamental operation in many scientific and engineering applications. With the introduction of matrix multiplication into Blosc2, users can now perform this operation on compressed arrays efficiently. The key advantages of having matrix multiplication in Blosc2 include:

• Compressed matrices in memory: Blosc2 enables matrices to be stored in a compressed format without sacrificing the ability to perform operations directly on them.

• Efficiency with chunks: In computation-intensive applications, matrix multiplication can be executed without fully decompressing the data, operating on small blocks of data independently, saving both time and memory.

• Out-of-core computation: When matrices are too large to fit in main memory, Blosc2 facilitates out-of-core processing. Data stored on disk is read and processed in optimized chunks, allowing matrix multiplication operations without loading the entire dataset into memory.

These features are especially valuable in big data environments and in scientific or engineering applications where matrix sizes can be overwhelming, enabling complex calculations efficiently.

Implementation

The matrix multiplication functionality is implemented in the matmul function. It supports Blosc2 NDArray objects and leverages chunked operations to perform the multiplication efficiently.

The image illustrates a blocked matrix multiplication approach. The key idea is to divide matrices into smaller blocks (or chunks) to optimize memory access and computational efficiency.

In the image, matrix A (M x K) and matrix B (K x N) are partitioned into chunks, and these are partitioned into blocks. The resulting matrix C (M x N) is computed as a sum of block-wise multiplication.

This method significantly improves cache utilization by ensuring that only the necessary parts of the matrices are loaded into memory at any given time. In Blosc2, storing matrix blocks as compressed chunks reduces memory footprint and enhances performance by enabling on-the-fly decompression.

Also, Blosc2 supports a wide range of data types. In addition to standard Python types such as int, float, and complex, it also fully supports various NumPy types. The currently supported types include:

• np.int8

• np.int16

• np.int32

• np.int64

• np.float32

• np.float64

• np.complex64

• np.complex128

This versatility allows compression and subsequent processing to be applied across diverse scenarios, tailored to the specific needs of each application.

Together, these features make Blosc2 a flexible and adaptable tool for various scenarios, but especially suited for the handling of large datasets.

Benchmarks

The benchmarks have been designed to evaluate the performance of the matmul function under various conditions. Here are the key aspects of our experimental setup and findings:

Different matrix sizes were tested using both float32 and float64 data types. All the matrices used for multiplication are square. The variation in matrix sizes helps observe how the function scales and how the overhead of chunk management impacts performance.

The x-axis represents the size of the resulting matrix in megabytes (MB). We used GFLOPS (Giga Floating-Point Operations per Second) to gauge the computational throughput, allowing us to compare the efficiency of the matmul function relative to highly optimized libraries like NumPy.

Blosc2 also incorporates a functionality to automatically select chunks, and it is represented in the benchmark by "Auto".

For smaller matrices, the overhead of managing chunks in Blosc2 can result in lower GFLOPS compared to NumPy. As the matrix size increases, Blosc2 scales well, approaching its performance to NumPy.

Each chunk shape exhibits a peak performance when the matrix size matches the chunk size, or is a multiple of the chunk shape.

Conclusion

The new matrix multiplication feature in Blosc2 introduces efficient, chunked computation for compressed arrays. This allows users to handle large datasets both in memory and on disk without sacrificing performance. The implementation supports a wide range of data types, making it versatile for various numerical applications.

Real-world applications, such as neural network training, demonstrate the potential benefits in scenarios where memory constraints and large data sizes are common. While there are some limitations —such as support only for 2D arrays and the overhead of blocking— the applicability looks promising, like potential integration with deep learning frameworks.

Overall, Blosc2 offers a compelling alternative for applications where the advantages of compression and out-of-core computation are critical, paving the way for more efficient processing of massive datasets.

Getting my feet wet with Blosc2

In the initial phase of the project, my biggest challenge was understanding how Blosc2 manages data internally. For matrix multiplication, it was critical to grasp how to choose the right chunks, since the operation requires that the ranges of both matrices coincide. After some considerations and a few insightful conversations with Francesc, I finally understood the underlying mechanics. This breakthrough allowed me to begin implementing the first versions of my solution, adjusting the data fragmentation so that each block was properly aligned for precise computation.

Another important aspect was adapting to the professional workflow of using Git for version control. Embracing Git —with its branch creation, regular commits, and conflict resolution— represented a significant shift in my development approach. This experience not only improved the organization of my code and facilitated collaboration but also instilled a structured and disciplined mindset in managing my projects. This tool has shown to be both valuable and extremely helpful.

Finally, the moment when the function finally returned the correct result was really exciting. After multiple iterations, the rigorous debugging process paid off as everything fell into place. This breakthrough validated the robustness of the implementation and boosted my confidence to further optimize and tackle new challenges in data processing.

Getting Started with Arrays and Broadcasting

Blosc2 works smoothly with arrays of various shapes and dimensions, enabling users to perform calculations such as addition or multiplication across arrays of different sizes. This is where broadcasting comes in. With broadcasting, Blosc2 automatically aligns the shapes of arrays for easy operations. This means you don’t need to manually adjust array dimensions to match, a huge time-saver when working with multidimensional data.

For example, let’s suppose we have an array representing a large dataset and, a, another representing a smaller dimension, c.

a = blosc2.full((1, 3, 2), fill_value=3)
c = blosc2.full(2, fill_value=9, dtype=np.int8)
expr = a + c - 1

As seen above, broadcasting works automatically (and efficiently) with arrays of compressed data. Also, the correct data type of the result will be inferred from the operands and the expression. Thanks to this mechanism, the interpreter automatically adjusts the dimensions and data types of the arrays involved in the operation, allowing calculations to be performed without the need for manual adjustments.

This approach is ideal for quick and simple data analysis, especially when working with large volumes of information that require frequent operations across different dimensions.

Setting Up and Saving Lazy Expressions

Imagine you need to perform a calculation like sum(a, axis=0) + b * sin(c). Rather than immediately calculating this, Blosc2’s lazy expression feature lets you store the expression for later. By using blosc2.lazyexpr, you define complex mathematical formulas and only trigger their execution when required, and only for the part of the resulting array that you are interested in. This is highly advantageous for large computations that might not be needed right away or that may depend on evolving data.

Let's see how that works with a little more complex expression:

# Create arrays with specific dimensions and values
a = blosc2.full((2, 3, 4), 1, urlpath="a.b2nd", mode="w")
b = blosc2.full((2, 4), 2, urlpath="b.b2nd", mode="w")
c = blosc2.full(4, 3, dtype=np.uint8, urlpath="c.b2nd", mode="w")
# Define a lazy expression and the operands for later execution
# Note that we are using a string version of the expression here
# so that it can be re-opened as-is later on
expression = "sum(a, axis=0) + b * sin(c)"
lazy_expression = blosc2.lazyexpr(expression)
lazy_expression.save("arrayResult.b2nd", mode="w")

In this code, sum(a, axis=0) + b * sin(c) is defined but not executed immediately. When you’re ready to use the result, you can call lazy_expression.compute() (returns a Blosc2 array that is compressed by default) to run the calculation. Alternatively, you can specify the part of the result that you are interested in with lazy_expression[0, :] (returns a NumPy array). This way, you save CPU and memory and only perform the computation when necessary.

Dynamic Computation: Reusing and Updating Results

Another big advantage of Blosc2 is its ability to compute persistent expressions that are dynamic: when an operand is enlarged, Blosc2 re-adapts the expression to account for its new shape. This approach significantly speeds up processing time, especially when working with frequently updated or real-time data.

For instance, if you have an expression stored, and only part of your dataset changes, Blosc2 can apply reductions dynamically to efficiently update the sum:

# Resizing arrays and updating values
a.resize((30, 30, 40))
a[20:30] = 5
b.resize((30, 40))
b[20:30] = 7
# Open the saved file
lazy_expression = blosc2.open(urlpath=url_path)
result = lazy_expression.compute()

In this case, the final result will have a shape of (30, 40) (instead of the previous (20, 40)). This allows for quick adaptability, which is crucial in data environments where values evolve constantly.

Why Persistent Reductions and Lazy Expressions Matter

These features make Blosc2 a top choice for working with large datasets, as they allow for:

• Broadcasting of memory, on-disk or network operands.

• Efficient use of CPU and memory by only executing calculations when needed.

• Dynamic expressions that adapt to changing data in operands.

• Enhanced performance due to streamlined, multi-threaded and pre-fetched calculations.

Together, lazy expressions and persistent reductions provide a flexible, resource-efficient way to manage complex data processes. They’re perfect for real-time analysis, evolving datasets, or any high-performance computing tasks requiring dynamic data handling.

Conclusion

Blosc2’s features offer a way to make data processing smarter and faster. If you work with large arrays or require adaptable workflows, Blosc2 can help you make the most of your data processing resources.

For more in-depth guidance, visit the full tutorial on Blosc2.

The 3D array

We will use a 3D array of type float64 with shape (1000, 1000, 1000). This array will be filled with values from 0 to 1000, and the goal will be to compute the sum of values in stripes of 100 elements in one axis, and including all the values in the other axis. We will perform reductions along the X, Y, and Z axes, comparing Blosc2 performance (with and without compression) against NumPy.

Reducing with NumPy

We will start by performing different sum reductions using NumPy. First, summing along the X, Y, and Z axes (and getting 2D arrays as result) and then summing along all axis (and getting an scalar as result).

axes = ("X", "Y", "Z", "all")
meas_np = {"sum": {}, "time": {}}
for n, axis in enumerate(axes):
    n = n if axis != "all" else None
    t0 = time()
    meas_np["sum"][axis] = np.sum(a, axis=n)
    t = time() - t0
    meas_np["time"][axis] = time() - t0

Reducing with Blosc2

Now let's create the Blosc2 array from the NumPy array. First, let's define the parameters for Blosc2: number of threads, compression levels, codecs, and chunk sizes. We will exercise different combinations of these parameters (including no compression) to evaluate the performance of Python-Blosc2 in reducing data in 3D arrays.

# Params for Blosc2
clevels = (0, 5)
codecs = (blosc2.Codec.LZ4, blosc2.Codec.ZSTD)

The function shown below is responsible for creating the different arrays and performing the reductions for each combination of parameters.

# Create a 3D array of type float64
def measure_blosc2(chunks):
    meas = {}
    for codec in codecs:
        meas[codec] = {}
        for clevel in clevels:
            meas[codec][clevel] = {"sum": {}, "time": {}}
            cparams = {"clevel": clevel, "codec": codec}
            a1 = blosc2.asarray(a, chunks=chunks, cparams=cparams)
            if clevel > 0:
                print(f"cratio for {codec.name} + SHUFFLE: {a1.schunk.cratio:.1f}x")
            # Iterate on Blosc2 and NumPy arrays
            for n, axis in enumerate(axes):
                n = n if axis != "all" else None
                t0 = time()
                # Perform the sum of the stripe (defined by the slice_)
                meas[codec][clevel]["sum"][axis] = a1.sum(axis=n)
                t = time() - t0
                meas[codec][clevel]["time"][axis] = t
                # If interested, you can uncomment the following line to check the results
                #np.testing.assert_allclose(meas[codec][clevel]["sum"][axis],
                #                           meas_np["sum"][axis])
    return meas

Automatic chunking

Let's plot the results for the X, Y, and Z axes, comparing the performance of Python-Blosc2 with different configurations against NumPy.

We can see that reduction along the X axis is much slower than those along the Y and Z axis for the Blosc2 case. This is because the automatically computed chunk shape is (1, 1000, 1000) making the overhead of partial sums larger. In addition, we see that, with the exception of the X axis, Blosc2+LZ4+SHUFFLE actually achieves far better performance than NumPy. Finally, when not using compression inside Blosc2, we never see an advantage. See later for a discussion on these results.

Manual chunking

Let's try to improve the performance by manually setting the chunk size. In the next case, we want to make performance similar along the three axes, so we will set the chunk size to (100, 100, 100) (8 MB).

In this case, performance in the X axis is already faster than Y and Z axes for Blosc2. Interestingly, performance is also faster than NumPy in X axis, while being very similar in Y and Z axis.

We could proceed further and try to fine tune the chunk size to get even better performance, but this is out of the scope of this blog (and more a task for Btune). Instead, we will try to make some sense on the results above; see below.

Why Blosc2 can be faster than NumPy?

As Blosc2 is using the NumPy machinery for computing reductions behind the scenes, why is Blosc2 faster than NumPy in several cases above? The answer lies in the way Blosc2 and NumPy access data in memory.

Blosc2 splits data into chunks and blocks to compress and decompress data efficiently. When accessing data, a full chunk is fetched from memory and decompressed by the CPU (as seen in the image below, left side). If the chunk size is small enough to fit in the CPU cache, the CPU can write the decompressed chunk faster, as it does not need to travel back to the main memory. Later, when NumPy is called to perform the reduction on the decompressed chunk, it can access the data faster, as it is already in the CPU cache (image below, right side).

But for allowing NumPy go faster, Blosc2 needs to decompress several chunks prior to NumPy performing the reduction operation. The decompressed chunks are stored on a queue, waiting for further processing; this is why Blosc2 needs to handle several (3 or 4) chunks simultaneously. In our case, the L3 cache size of our CPU (Intel 13900K) is 36 MB, and Blosc2 has chosen 8 MB for the chunk size, allowing to store up to 4 chunks in L3, which is near to optimal. Also, when we have chosen the chunk size to be (100, 100, 100), the chunk size is still 8 MB, which continues to be fine indeed.

All in all, it is not that Blosc2 is faster than NumPy, but rather that it is allowing NumPy to leverage the CPU cache more efficiently. Having said this, we still need some explanation on why the performance can be so different along the X, Y, and Z axes, specially for the first chunk shape (automatic) above. Let's address this in the next section.

Performing reductions on 3D arrays

On a three-dimensional environment, like the one shown in the image, data is organized in a cubic space with three axes: X, Y, and Z. By default, Blosc2 chooses the chunk size so that it fits in the CPU cache comfortably. On the other hand, it tries to follow the NumPy convention of storing data row-wise; so, this is why the default chunk shape has been chosen as (1, 1000, 1000). In this case, it is clear that reduction times along different axes are not going to be the same, as the sizes of the chunk in different axes are not uniform (actually, there is a large asymmetry).

The difference in cost while traversing data values can be visualized more easily on a 2D array:

Reduction along the X axis: When accessing a row (red line), the CPU can access these values (red points) from memory sequentially, but they need to be stored on an accumulator. The next rows needs to be fetched from memory and be added to the accumulator. If the size of the accumulator is large (in this case is 1000 * 1000 * 8 = 8 MB), it does not fit in low level CPU caches, and has to be peformed in the relatively slow L3.

Reducing along the Y axis: When accessing a row (green line), the CPU can access these values (green points) from memory sequentially but, contrarily to the case above, they don't even need an accumulator, and the sum of the row (marked as an *) is final. So, although the number of sum operations is the same as above, the required time is smaller because there is no need of updating all the values of the accumulator per row, but only one at a time, which is faster in modern CPUs.

Tweaking the chunk size

However, when Blosc2 is instructed to create chunks that are the same size for all the axes (chunks=(100, 100, 100)), the situation changes. In this case, an accumulator is needed for each chunk (sub-cube in figure above), but as it is relatively small (100 * 100 * 8 = 80 KB), and fits in L2, so accumulation in the X axis is faster than in the previous scenario (remember that it needs to do the accumulation in L3).

Incidentally, now Blosc2 performance along X axis is even better than in the Y and Z axes, as the CPU can access data in a more efficient way. Furthermore, Blosc2 performance is up to 1.5x better than NumPy in the X axis (while being similar, or even a bit better along Y and Z axes), which is a quite remarkable feat.

Conclusion

Understanding the balance between space savings and the additional time required to process the data is important. Testing different compression settings can help finding the method that offers the best trade-off between reduced size and processing time. The fact that Blosc2 automatically chooses the chunk shape, makes it easy for the user to get a decently good performance, without having to worry about the details of the CPU cache. In addition, as we have shown, we can fine tune the chunk shape in case the default one does not fit our needs (e.g. we need more uniform performance along all axes).

Besides the sum() reduction exercised here, Blosc2 supports a fair range of reduction operators (mean, std, min, max, all, any, etc.), and you are invited to explore them. Moreover, it is also possible to use reductions even for very large arrays that are stored on disk. This opens the door to a wide range of possibilities for data analysis and science, allowing for efficient reductions on large datasets that are compressed on-disk and with minimal memory usage. We will explore this in a forthcoming blog.

We would like to thank ironArray for supporting the development of the computing capabilities of Blosc2. Then, to NumFOCUS for recently providing a small grant that is helping us to improve the documentation for the project. Last but not least, we would like to thank the Blosc community for providing so many valuable insights and feedback that have helped us to improve the performance and usability of Blosc2.

Using the API

The direct chunking API consists of three operations: get information about a chunk (chunk_info()), write a raw chunk (write_chunk()), and read a raw chunk (read_chunk()). They are supported by chunked datasets (CArray, EArray and Table), i.e. those whose data is split into fixed-size chunks of the same dimensionality as the dataset (maybe padded at its boundaries), with HDF5 pipeline filters like compressors optionally processing them on read/write.

chunk_info() returns an object with useful information about the chunk containing the item at the given coordinates. Let's create a simple 100x100 array with 10x100 chunks compressed with Blosc2+LZ4 and get info about a chunk:

>>> import tables, numpy
>>> h5f = tables.open_file('direct-example.h5', mode='w')
>>> filters = tables.Filters(complib='blosc2:lz4', complevel=2)
>>> data = numpy.arange(100 * 100).reshape((100, 100))
>>> carray = h5f.create_carray('/', 'carray', chunkshape=(10, 100),
                               obj=data, filters=filters)
>>> coords = (42, 23)
>>> cinfo = carray.chunk_info(coords)
>>> cinfo
ChunkInfo(start=(40, 0), filter_mask=0, offset=6779, size=608)

So the item at coordinates (42, 23) is stored in a chunk of 608 bytes (compressed) which starts at coordinates (40, 0) in the array and byte 6779 in the file. The latter offset may be used to let other code access the chunk directly on storage. For instance, since Blosc2 was the only HDF5 filter used to process the chunk, let's open it directly:

>>> import blosc2
>>> h5f.flush()
>>> b2chunk = blosc2.open(h5f.filename, mode='r', offset=cinfo.offset)
>>> b2chunk.shape, b2chunk.dtype, data.itemsize
((10, 100), dtype('V8'), 8)

Since Blosc2 does understand the structure of data (thanks to b2nd), we can even see that the chunk shape and the data item size are correct. The data type is opaque to the HDF5 filter which wrote the chunk, hence the V8 dtype. Let's check that the item at (42, 23) is indeed in that chunk:

>>> chunk = numpy.ndarray(b2chunk.shape, buffer=b2chunk[:],
                          dtype=data.dtype)  # Use the right type.
>>> ccoords = tuple(numpy.subtract(coords, cinfo.start))
>>> bool(data[coords] == chunk[ccoords])
True

This offset-based access is actually what b2nd optimized slicing performs internally. Please note that neither PyTables nor HDF5 were involved at all in the actual reading of the chunk (Blosc2 just got a file name and an offset). It's difficult to cut more overhead than that!

It won't always be the case that you can (or want to) read a chunk in that way. The read_chunk() method allows you to read a raw chunk as a new byte string or into an existing buffer, given the chunk's start coordinates (which you may compute yourself or get via chunk_info()). Let's use read_chunk() to redo the reading that we just did above:

>>> rchunk = carray.read_chunk(coords)
Traceback (most recent call last):
    ...
tables.exceptions.NotChunkAlignedError: Coordinates are not multiples
    of chunk shape: (42, 23) !* (np.int64(10), np.int64(100))
>>> rchunk = carray.read_chunk(cinfo.start)  # Always use chunk start!
>>> b2chunk = blosc2.ndarray_from_cframe(rchunk)
>>> chunk = numpy.ndarray(b2chunk.shape, buffer=b2chunk[:],
                          dtype=data.dtype)  # Use the right type.
>>> bool(data[coords] == chunk[ccoords])
True

The write_chunk() method allows you to write a byte string into a raw chunk. Please note that you must first apply any filters manually, and that you can't write chunks beyond the dataset's current shape. However, remember that enlargeable datasets may be grown or shrunk in an efficient manner using the truncate() method, which doesn't write new chunk data. Let's use that to create an EArray with the same data as the previous CArray, chunk by chunk:

>>> earray = h5f.create_earray('/', 'earray', chunkshape=carray.chunkshape,
                               atom=carray.atom, shape=(0, 100),  # Empty.
                               filters=filters)  # Just to hint readers.
>>> earray.write_chunk((0, 0), b'whatever')
Traceback (most recent call last):
    ...
IndexError: Chunk coordinates not within dataset shape:
    (0, 0) <> (np.int64(0), np.int64(100))
>>> earray.truncate(len(carray))  # Grow the array (cheaply) first!
>>> for cstart in range(0, len(carray), carray.chunkshape[0]):
...     chunk = carray[cstart:cstart + carray.chunkshape[0]]
...     b2chunk = blosc2.asarray(chunk)  # May be customized.
...     wchunk = b2chunk.to_cframe()  # Serialize.
...     earray.write_chunk((cstart, 0), wchunk)

You can see that such low-level writing is more involved than usual. Though we used default Blosc2 parameters here, the explicit compression step allows you to fine-tune it in ways not available through PyTables like setting internal chunk and block sizes or even using Blosc2 compression plugins like Grok/JPEG2000. In fact, the filters given on dataset creation are only used as a hint, since each Blosc2 container holding a chunk includes enough metadata to process it independently. In the example, the default chunk compression parameters don't even match dataset filters (using Zstd instead of LZ4):

>>> carray.filters
Filters(complevel=2, complib='blosc2:lz4', ...)
>>> earray.filters
Filters(complevel=2, complib='blosc2:lz4', ...)
>>> b2chunk.schunk.cparams['codec']
<Codec.ZSTD: 5>

Still, the Blosc2 HDF5 filter plugin included with PyTables is able to read the data just fine:

>>> bool((carray[:] == earray[:]).all())
True
>>> h5f.close()

You may find a more elaborate example of using direct chunking in PyTables' examples.

Benchmarks

b2nd optimized slicing shows us that removing the HDF5 filter pipeline from the I/O path can result in sizable performance increases, if the right chunking and compression parameters are chosen. To check the impact of using the new direct chunking API, we ran some benchmarks that compare regular and direct read/write speeds. On an AMD Ryzen 7 7800X3D CPU with 8 cores, 96 MB L3 cache and 8 MB L2 cache, clocked at 4.2 GHz, we got the following results:

We can see that direct chunking yields 3.75x write and 4.4x read speedups, reaching write/read speeds of 1.7 GB/s and 5.2 GB/s. These are quite impressive numbers, though the base equipment is already quite powerful. Thus we also tried the same benchmark on a consumer-level MacBook Air laptop with an Apple M1 CPU with 4+4 cores and 12 MB L2 cache, clocked at 3.2 GHz, with the following results:

In this case direct chunking yields 4.5x write and 1.9x read speedups, with write/read speeds of 0.8 GB/s and 1.6 GB/s. The absolute numbers are of course not as impressive, but the performance is still much better than that of the regular mechanism, especially when writing. Please note that the M1 CPU has a hybrid efficiency+performance core configuration; as an aside, running the benchmark on a high-range Intel Core i9-13900K CPU also with a hybrid 8+16 core configuration (32 MB L2, 5.7 GHz) raised the write speedup to 4.6x, reaching an awesome write speed of 2.6 GB/s.

All in all, it's clear that bypassing the HDF5 filter pipeline results in immediate I/O speedups. You may find a Jupyter notebook with the benchmark code and AMD CPU data in PyTables' benchmarks.

Conclusions

First of all, we (Ivan Vilata and Francesc Alted) want to thank everyone who made this new 3.10 release of PyTables possible, especially Antonio Valentino for his role of project maintainer, and the many code and issue contributors. Trying the new direct chunking API is much easier because of them. And of course, a big thank you to the NumFOCUS Foundation for making this whole new feature possible by funding its development!

In this post we saw how PyTables' direct chunking API allows one to squeeze the extra drop of performance that the most demanding scenarios require, when adjusting chunking and compression parameters in PyTables itself can't go any further. Of course, its low-level nature makes its use less convenient and safe than higher-level mechanisms, so you should evaluate whether the extra effort pays off. If used carefully with robust filters like Blosc2, the direct chunking API should shine the most in the case of large datasets with sustained I/O throughput demands, while retaining compatibility with other HDF5-based tools.

Understanding lossy compression

Unlike lossless compression, where the original data can be perfectly reconstructed from the compressed version, lossy compression involves discarding some information to achieve higher compression ratios. While this inevitably results in a loss of fidelity, the trade-off is often justified by the significant reduction in storage size.

Lossy compression techniques are commonly employed in scenarios where minor degradation in quality is acceptable, such as multimedia applications (e.g., images, audio, and video) and scientific data analysis. By intelligently discarding less crucial information, lossy compression algorithms can achieve substantial compression ratios while maintaining perceptual quality within acceptable bounds.

Lossy codecs in Blosc2

In the context of Blosc2, lossy compression can be achieved either through a combination of traditional compression algorithms and filters that can selectively discard less critical data, or by using codecs specially meant for doing so.

Applications and use cases

The versatility of Blosc2's lossy compression capabilities opens up a myriad of applications across different domains. In scientific computing, for example, where large volumes of data are generated and analyzed, lossy compression can significantly reduce storage requirements without significantly impacting the accuracy of results.

Similarly, in multimedia applications, such as image and video processing, lossy compression can help minimize bandwidth usage and storage costs while maintaining perceptual quality within acceptable limits.

Compressing images with JPEG 2000 and with INT_TRUNC

As an illustration, a recent study involved the compression of substantial volumes of 16-bit grayscale images sourced from different synchrotron facilities in Europe. While achieving efficient compression ratios necessitates the use of lossy compression techniques, it is essential to exercise caution to preserve key features for clear visual examination and accurate numerical analysis. Below, we provide an overview of how Blosc2 can employ various codecs and quality settings within filters to accomplish this task.

The SSIM index, derived from the Structural Similarity Measure, gauges the perceived quality of an image, with values closer to 1 indicating higher fidelity. You can appreciate the varying levels of fidelity achievable through the utilization of different filters and codecs.

In terms of performance, each of these compression methods also showcases significantly varied speeds (tested on a MacBook Air with an M1 processor):

A pivotal benefit of Blosc2's strategy for lossy compression lies in its adaptability and configurability. This enables tailoring to unique needs and limitations, guaranteeing optimal performance across various scenarios.

Conclusion

Lossy compression is a powerful tool for optimizing storage space, reducing bandwidth usage, and improving overall efficiency in data handling. With Blosc2, developers have access to a robust and flexible compression library for both lossless and lossy compression modes.

With its advanced compression methodologies and adept memory management, Blosc2 empowers users to strike a harmonious balance between compression ratio, speed, and fidelity. This attribute renders it especially suitable for scenarios where resource limitations or performance considerations hold significant weight.

Finally, there are ongoing efforts towards integrating fidelity into our BTune AI tool. This enhancement will empower the tool to autonomously identify the most suitable codecs and filters, balancing compression level, precision, and fidelity according to user-defined preferences. Keep an eye out for updates!

Whether you're working with scientific data, multimedia content, or large-scale datasets, Blosc2 offers a comprehensive solution for efficient data compression and handling.

Why JPEG2000?

It can be stated that currently the best compromise between image quality and compression factor is JPEG2000. This image compression standard has been in use for over 20 years and is widely accepted in the imaging community due to its ability to achieve a compression factor of approximately 10x without significant loss of information.

JPEG2000 has been implemented in many libraries, but the one that stands out is the grok library because of its speed and completeness. This is why we have chosen it to be the first image codec to be added to Blosc2.

Installing the plugin

First of all, you will need to install the blosc2-grok plugin. You can do it with:

pip install blosc2-grok

That's it! You are ready to use it.

Using the grok codec via the plugin

The grok codec has been already registered as a global plugin for Blosc2, so you only need to use its plugin id (blosc2.Codec.GROK) in the codec field of the cparams:

# Define the compression parameters. Disable the filters and the
# splitmode, because these don't work with the codec.
cparams = {
    'codec': blosc2.Codec.GROK,
    'filters': [],
    'splitmode': blosc2.SplitMode.NEVER_SPLIT,
}

It is important to disable any filter or splitmode, since we don't want the data to be modified before proceeding to the compression using grok.

Now, imagine you have an image as a NumPy array (let's say, created using pillow), and you want to compress via blosc2_grok. Before, you will need to tell blosc2-grok which format to use among the available ones in the grok library (we will get through the different parameters later):

# Set the parameters that will be used by the codec
kwargs = {'cod_format': blosc2_grok.GrkFileFmt.GRK_FMT_JP2}
blosc2_grok.set_params_defaults(**kwargs)

And finally, you are able to compress the image with:

bl_array = blosc2.asarray(
    np_array,
    chunks=np_array.shape,
    blocks=np_array.shape,
    cparams=cparams,
    urlpath="myfile.b2nd",
    mode="w",
)

We already have compressed our first image with blosc2-grok!

In this case, the chunks and blocks params of Blosc2 have been set to the shape of the image (including the number of components) so that grok receives the image as a whole and therefore, can find more opportunities to compress better.

Setting grok parameters

We have already used the cod_format grok parameter to set the format to use with the blosc2_grok.set_params_defaults(), but blosc2-grok lets you set many other parameters. All of them are mentioned in the README. When possible, and to make it easier for existing users, these parameters are named the same than in the Pillow library.

For example, let's see how to set the quality_mode and quality_layers params, which are meant for lossy compression. So, realize you don't care too much about the quality but want a compression ratio of 10x. Then, you would specify the quality_mode to be rates and the quality_layers to 10:

kwargs = {'cod_format': blosc2_grok.GrkFileFmt.GRK_FMT_JP2}
kwargs['quality_mode'] = 'rates'
kwargs['quality_layers'] = np.array([10], dtype=np.float64)
blosc2_grok.set_params_defaults(**kwargs)

With that, you will be able to store the same image than before, but with a compression ratio of 10x. Please note that the quality_layers parameter is a numpy array. By the way, specifying more than one element here will produce different layers of quality of the original image, but this has little use in Blosc2, since it is better to store different layers in different NDArray objects (or files).

Now, just like in Pillow, quality_mode can also be expressed in dB, which indicates that you want to specify the quality as the peak signal-to-noise ratio (PSNR) in decibels. For example, let's set a PSNR of 45 dB (which will give us a compression of 9x):

kwargs['quality_mode'] = 'dB'
kwargs['quality_layers'] = np.array([45], dtype=np.float64)

Another useful parameter if you want to speed things up is the num_threads parameter. Although grok already sets a good default for you, you can set it to some other value (e.g. when experimenting the best one for your box). Or, if you would like to deactivate multithreading, you can set it to 1:

kwargs['num_threads'] = 1

For example, in a MacBook Air laptop with Apple M2 CPU (8-core), the speed difference when performing lossless compression between the single thread setting and the default thread value is around 6x, so expect quite large accelerations by leveraging multithreading.

Visual example

Below we did a total of 3 different compressions. First an image using lossless compression showing the original image:

Then, a couple of images using lossy compression: one with 10x for rates quality mode (left) and another with 45dB for dB quality mode (right):

As can be seen, the lossy images have lost some quality which is to be expected when using this level of compression (around 10x), but the great quality of the JPEG2000 codec allows us human beings to still perceive the image quite well.

A glimpse on performance

The combination of the great implementation of the JPEG2000 codec in grok and the multithreading capabilities of Blosc2 allow to compress, but specially decompress, the image very fast (benchmark run on an Intel i9-13900K CPU):

One can see that the compression speed is quite good (around 140 MB/s), but that the decompression speed is much faster (up to 800 MB/s). See how, in comparison, the compression speed of the JPEG2000 in Pillow (via the OpenJPEG codec) is much slower (around 4.5 MB/s max.) and so is the decompression speed (around 16 MB/s max.).

Besides, both grok and OpenJPEG can achieve very similar quality when using similar compression ratios. For example, when using the structural similarity index measure (SSIM) to compare the original image with the decompressed one, we get the following results:

Actually, the flexibility of the double partitioning in Blosc2 allows for quite a few ways to divide the workload during compression/decompression, affecting both speed and quality, but we will leave this discussion for another blog. If you are interested in this topic, you can have a look at the blosc2-grok benchmarks.

Conclusion

The addition of the grok plugin to Blosc2 opens many possibilities for compressing images. In the example we used a RGB image, but grayscale images, up to 16-bit of precision, can also be compressed without any problem.

Although fully usable, this plugin is still in its early stages, so we encourage you to try it out and give us feedback; we will be happy to hear from you!

Thanks to the LEAPS consortium for sponsoring this work, and NumFOCUS for their continuous support through the past years. Thanks also to Aaron Boxer, for the excellent grok library, and for his help in making this plugin possible. If you like what we are doing in the Blosc world, please consider donating to the project.

Direct chunk access and two-level partitioning

You may remember that the previous version of PyTables (3.8.0) already got support for direct access to Blosc2-compressed chunks (bypassing the HDF5 filter pipeline), with two-level partitioning of chunks into smaller blocks (allowing for fast access to small slices with big chunks). You may want to read Óscar and Francesc's post Blosc2 Meets PyTables to see the great performance gains provided by these techniques.

However, these enhancements only applied to tabular datasets, i.e. one-dimensional arrays of a uniform, fixed set of fields (columns) with heterogeneous data types as illustrated above. Multi-dimensional compressed arrays of homogeneous data (another popular feature of PyTables) still used plain chunking going through the HDF5 filter pipeline, and flat chunk compression. Thus, they suffered from the high overhead of the very generic pipeline and the inefficient decompression of whole (maybe big) chunks even for small slices.

Now, you may have also read the post by the Blosc Development Team about Blosc2 NDim (b2nd), first included in C-Blosc 2.7.0. It introduces the generalization of Blosc2's two-level partitioning to multi-dimensional arrays as shown below. This makes arbitrary slicing of such arrays across any dimension very efficient (as better explained in the post about Caterva, the origin of b2nd), when the right chunk and block sizes are chosen.

This b2nd support was the missing piece to extend PyTables' chunking and slicing optimizations from tables to uniform arrays.

Choosing adequate chunk and block sizes

Let us try a benchmark very similar to the one in the post introducing Blosc2 NDim, which slices a 50x100x300x250 floating-point array (2.8 GB) along its four dimensions, but this time with 64-bit integers, and using PyTables 3.9 with flat slicing (via the HDF5 filter pipeline), PyTables 3.9 with b2nd slicing (optimized, via direct chunk access implemented in C), h5py 3.10 with flat slicing (via hdf5plugin 4.3's support for Blosc2 in the HDF5 filter pipeline), and h5py with b2nd slicing (via the experimental b2h5py package using direct chunk access implemented in Python through h5py).

According to the aforementioned post, Blosc2 works better when blocks have a size which allows them to fit both compressed and uncompressed in each CPU core’s L2 cache. This of course depends on the data itself and the compression algorithm and parameters chosen. Let us choose LZ4+shuffle since it offers a reasonable speed/size trade-off, and try to find the different compression levels that work well with our CPU (level 8 seems best in our case).

With the benchmark's default 10x25x50x50 chunk shape, and after experimenting with the BLOSC_NTHREADS environment variable to find the number of threads that better exploit Blosc2's parallelism (6 for our CPU), we obtain the results shown below:

The optimized b2nd slicing of PyTables already provides some speedups (although not that impressive) in the inner dimensions, in comparison with flat slicing based on the HDF5 filter pipeline (which performs similarly for PyTables and h5py). As explained in Blosc2 Meets PyTables, HDF5 handling of chunked datasets favours big chunks that reduce in-memory structures, while Blosc2 can further exploit parallel threads to handle the increased number of blocks. Our CPU's L3 cache is 36MB big, so we may still grow the chunksize to reduce HDF5 overhead (without hurting Blosc2 parallelism).

Let us raise the chunkshape to 10x25x150x100 (28.6MB) and repeat the benchmark (again with 6 Blosc2 threads):

Much better! Choosing a better chunkshape not just provides up to 10x speedup for the PyTables optimized case, it also results in 4x-5x speedups compared to the performance of the HDF5 filter pipeline. The optimizations applied to h5py also yield considerable speedups (for an initial, Python-based implementation).

Conclusions and future work

The benchmarks above show how optimized Blosc2 NDim's two-level partitioning combined with direct HDF5 chunk access can yield considerable performance increases when slicing multi-dimensional Blosc2-compressed arrays under PyTables (and h5py). However, the usual advice holds to invest some effort into fine-tuning some of the parameters used for compression and chunking for better results. We hope that this article also helps readers find those parameters.

It is worth noting that these techniques still have some limitations: they only work with contiguous slices (that is, with step 1 on every dimension), and on datasets with the same byte ordering as the host machine. Also, although results are good indeed, there may still be room for implementation improvement, but that will require extra code profiling and parameter adjustments.

Finally, as mentioned in the Blosc2 NDim post, if you need help in finding the best parameters for your use case, feel free to reach out to the Blosc team at contact (at) blosc.org.

Enjoy data!

Creating a dynamically loaded filter

To learn how to create dynamic plugins, we'll use an already created example. Suppose you have a filter that you want Blosc2 to load dynamically only when it is used. In this case, you need to create a Python package to build a wheel and install it as a separate library. You can follow the structure used in blosc2_plugin_example to do this:

├── CMakeLists.txt
├── README.md
├── blosc2_plugin_name
│   └── __init__.py
├── examples
│   ├── array_roundtrip.py
│   ├── schunk_roundtrip.py
│   └── test_plugin.c
├── pyproject.toml
├── requirements-build.txt
├── setup.py
└── src
    ├── CMakeLists.txt
    ├── urfilters.c
    └── urfilters.h

Note that the project name has to be blosc2_ followed by the plugin name (plugin_example in this case). The corresponding functions will be defined in the src folder, in our case in urfilters.c, following the same format as functions for user-defined filters (see https://github.com/Blosc/c-blosc2/blob/main/plugins/README.md for more information). Here it is the sample code:

int blosc2_plugin_example_forward(const uint8_t* src, uint8_t* dest,
                                  int32_t size, uint8_t meta,
                                  blosc2_cparams *cparams, uint8_t id) {
  blosc2_schunk *schunk = cparams->schunk;

  for (int i = 0; i < size / schunk->typesize; ++i) {
    switch (schunk->typesize) {
      case 8:
        ((int64_t *) dest)[i] = ((int64_t *) src)[i] + 1;
        break;
      default:
        BLOSC_TRACE_ERROR("Item size %d not supported", schunk->typesize);
        return BLOSC2_ERROR_FAILURE;
    }
  }
  return BLOSC2_ERROR_SUCCESS;
}


int blosc2_plugin_example_backward(const uint8_t* src, uint8_t* dest, int32_t size,
                                   uint8_t meta, blosc2_dparams *dparams, uint8_t id) {
  blosc2_schunk *schunk = dparams->schunk;

  for (int i = 0; i < size / schunk->typesize; ++i) {
    switch (schunk->typesize) {
      case 8:
        ((int64_t *) dest)[i] = ((int64_t *) src)[i] - 1;
        break;
      default:
        BLOSC_TRACE_ERROR("Item size %d not supported", schunk->typesize);
        return BLOSC2_ERROR_FAILURE;
    }
  }
  return BLOSC2_ERROR_SUCCESS;
}

In addition to these functions, we need to create a filter_info (or codec_info or tune_info in each case) named info. This variable will contain the names of the forward and backward functions. In our case, we will have:

filter_info info  = {"blosc2_plugin_example_forward", "blosc2_plugin_example_backward"};

To find the functions, the variable must always be named info. Furthermore, the symbols info and the functions forward and backward must be exported in order for Windows to find them. You can see all the details for doing that in the blosc2_plugin_example repository.

Creating and installing the wheel

Once the project is done, you can create a wheel and install it locally:

python setup.py bdist_wheel
pip install dist/*.whl

This wheel can be uploaded to PyPI so that anybody can use it. Once tested and stable enough, you can request the Blosc Team to register it globally. This way, an ID for the filter or codec will be booked so that the data will always be able to be encoded/decoded by the same code, ensuring portability.

Registering the plugin in C-Blosc2

After installation, and prior to use it, you must register it in C-Blosc2. This step is necessary only if the filter is not already registered globally by C-Blosc2, which is likely if you are testing it or you are not ready to share it with other users. To register it, follow the same process as registering a user-defined plugin, but leave the function pointers as NULL:

blosc2_filter plugin_example;
plugin_example.id = 250;
plugin_example.name = "plugin_example";
plugin_example.version = 1;
plugin_example.forward = NULL;
plugin_example.backward = NULL;
blosc2_register_filter(&plugin_example);

When the filter is used for the first time, C-Blosc2 will automatically fill in the function pointers.

Registering the plugin in Python-Blosc2

The same applies for Python-Blosc2. You can register the filter as follows:

import blosc2
blosc2.register_filter(250, None, None, "plugin_example")

Using the plugin in C-Blosc2

To use the plugin, simply set the filter ID in the filters pipeline, as you would do with user-defined filters:

blosc2_cparams cparams = BLOSC2_CPARAMS_DEFAULTS;
cparams.filters[4] = 250;
cparams.filters_meta[4] = 0;

blosc2_dparams dparams = BLOSC2_DPARAMS_DEFAULTS;

blosc2_schunk* schunk;

/* Create a super-chunk container */
cparams.typesize = sizeof(int32_t);
blosc2_storage storage = {.cparams=&cparams, .dparams=&dparams};
schunk = blosc2_schunk_new(&storage);

To see a full usage example, refer to https://github.com/Blosc/blosc2_plugin_example/blob/main/examples/test_plugin.c. Keep in mind that the executable using the plugin must be launched from the same virtual environment where the plugin wheel was installed. When compressing or decompressing, C-Blosc2 will dynamically load the library and call its functions automatically (as depicted below).

Once you are satisfied with your plugin, you may choose to request the Blosc Development Team to register it as a global plugin. The only difference (aside from its ID number) is that users won't need to register it locally anymore. Also, a dynamic plugin will not be loaded until it is explicitly requested by any compression or decompression function, saving resources.

Using the plugin in Python-Blosc2

As in C-Blosc2, just set the filter ID in the filters pipeline, as you would do with user-defined filters:

shape = [100, 100]
size = int(np.prod(shape))
nparray = np.arange(size, dtype=np.int32).reshape(shape)
blosc2_array = blosc2.asarray(nparray, cparams={"filters": [250]})

To see a full usage example, refer to https://github.com/Blosc/blosc2_plugin_example/blob/main/examples/array_roundtrip.py.

Conclusions

C-Blosc2's ability to support dynamically loaded plugins allows the library to grow in features without increasing the size and complexity of the library itself. For more information about user-defined plugins, refer to this blog entry.

We appreciate your interest in our project! If you find our work useful and valuable, we would be grateful if you could support us by making a donation. Your contribution will help us continue to develop and improve Blosc packages, making them more accessible and useful for everyone. Our team is committed to creating high-quality and efficient software, and your support will help us to achieve this goal.

Compressing ERA5 datasets

The best approach to evaluate a new filter is to apply it to real data. For this, we will use some of the ERA5 datasets, representing different measurements and labeled as "wind", "snow", "flux", "pressure" and "precip". They all contain floating point data (float32) and we will use a full month of each one, accounting for 2.8 GB for each dataset.

The diverse datasets exhibit rather dissimilar complexity, which proves advantageous for testing diverse compression scenarios. For instance, the wind dataset appears as follows:

The image shows the intricate network of winds across the globe on October 1, 1987. The South American continent is visible on the right side of the map.

Another example is the snow dataset:

This time the image is quite flat. Here one can spot Antarctica, Greenland, North America and of course, Siberia, which was pretty full of snow by 1987-10-01 23:00:00 already.

Let's see how the new bytedelta filter performs when compressing these datasets. All the plots below have been made using a box with an Intel i13900k processor, 32 GB of RAM and using Clear Linux.

In the box plot above, we summarized the compression ratios for all datasets using different codecs (BLOSCLZ, LZ4, LZ4HC and ZSTD). The main takeaway is that using bytedelta yields the best median compression ratio: bytedelta achieves a median of 5.86x, compared to 5.62x for bitshuffle, 5.1x for shuffle, and 3.86x for codecs without filters. Overall, bytedelta seems to improve compression ratios here, which is good news.

While the compression ratio is a useful metric for evaluating the new bytedelta filter, there is more to consider. For instance, does the filter work better on some data sets than others? How does it impact the performance of different codecs? If you're interested in learning more, read on.

Effects on various datasets

Let's see how different filters behave on various datasets:

Here we see that, for datasets that compress easily (precip, snow), the behavior is quite different from those that are less compressible. For precip, bytedelta actually worsens results, whereas for snow, it slightly improves them. For less compressible datasets, the trend is more apparent, as can be seen in this zoomed in image:

In these cases, bytedelta clearly provides a better compression ratio, most specifically with the pressure dataset, where compression ratio by using bytedelta has increased by 25% compared to the second best, bitshuffle (5.0x vs 4.0x, using ZSTD clevel 9). Overall, only one dataset (precip) shows an actual decrease. This is good news for bytedelta indeed.

Furthermore, Blosc2 supports another compression parameter for splitting the compressed streams into bytes with the same significance. Normally, this leads to better speed but less compression ratio, so this is automatically activated for faster codecs, whereas it is disabled for slower ones. However, it turns out that, when we activate splitting for all the codecs, we find a welcome surprise: bytedelta enables ZSTD to find significantly better compression paths, resulting in higher compression ratios.

As can be seen, in general ZSTD + bytedelta can compress these datasets better. For the pressure dataset in particular, it goes up to 5.7x, 37% more than the second best, bitshuffle (5.7x vs 4.1x, using ZSTD clevel 9). Note also that this new highest is 14% more than without splitting (the default).

This shows that when compressing, you cannot just trust your intuition for setting compression parameters - there is no substitute for experimentation.

Effects on different codecs

Now, let's see how bytedelta affects performance for different codecs and compression levels.

Interestingly, on average bytedelta proves most useful for ZSTD and higher compression levels of ZLIB (Blosc2 comes with ZLIB-NG). On the other hand, the fastest codecs (LZ4, BLOSCLZ) seem to benefit more from bitshuffle instead.

Regarding compression speed, in general we can see that bytedelta has little effect on performance:

As we can see, compression algorithms like BLOSCLZ, LZ4 and ZSTD can achieve extremely high speeds. LZ4 reaches and surpasses speeds of 30 GB/s, even when using bytedelta. BLOSCLZ and ZSTD can also exceed 20 GB/s, which is quite impressive.

Let’s see the compression speed grouped by compression levels:

Here one can see that, to achieve the highest compression rates when combined with shuffle and bytedelta, the codecs require significant CPU resources; this is especially noticeable in the zoomed-in view:

where capable compressors like ZSTD do require up to 2x more time to compress when using bytedelta, especially for high compression levels (6 and 9).

Now, let us examine decompression speeds:

In general, decompression is faster than compression. BLOSCLZ, LZ4 and LZ4HC can achieve over 100 GB/s. BLOSCLZ reaches nearly 180 GB/s using no filters on the snow dataset (lowest complexity).

Let’s see the decompression speed grouped by compression levels:

The bytedelta filter noticeably reduces speed for most codecs, up to 20% or more. ZSTD performance is less impacted.

Achieving a balance between compression ratio and speed

Often, you want to achieve a good balance of compression and speed, rather than extreme values of either. We will conclude by showing plots depicting a combination of both metrics and how bytedelta influences them.

Let's first represent the compression ratio versus compression speed:

As we can see, the shuffle filter is typically found on the Pareto frontier (in this case, the point furthest to the right and top). Bytedelta comes next. In contrast, not using a filter at all is on the opposite side. This is typically the case for most real-world numerical datasets.

Let's now group filters and datasets and calculate the mean values of combining (in this case, multiplying) the compression ratio and compression speed for all codecs.

As can be seen, bytedelta works best with the wind dataset (which is quite complex), while bitshuffle does a good job in general for the others. The shuffle filter wins on the snow dataset (low complexity).

If we group by compression level:

We see that bytedelta works well with LZ4 here, and also with ZSTD at the lowest compression level (1).

Let's revise the compression ratio versus decompression speed comparison:

Let's group together the datasets and calculate the mean for all codecs:

In this case, shuffle generally prevails, with bitshuffle also doing reasonably well, winning on precip and pressure datasets.

Also, let’s group the data by compression level:

We find that bytedelta compression does not outperform shuffle compression in any scenario. This is unsurprising since decompression is typically fast, and bytedelta's extra processing can decrease performance more easily. We also see that LZ4HC (clevel 6 and 9) + shuffle strikes the best balance in this scenario.

Finally, let's consider the balance between compression ratio, compression speed, and decompression speed:

Here the winners are shuffle and bitshuffle, depending on the data set, but bytedelta never wins.

If we group by compression levels:

Overall, we see LZ4 as the clear winner at any level, especially when combined with shuffle. On the other hand, bytedelta did not win in any scenario here.

Benchmarks for other computers

We have run the benchmarks presented here in an assortment of different boxes:

• MacBook Air with M1 processor and 8 GB RAM. MacOSX 13.1.

• AMD Ryzen 9 5950X processor and 32 GB RAM. Debian 22.04.

• Intel i9-10940X processor and 64 GB RAM. Debian 22.04.

• Intel i9-13900K processor and 32 GB RAM. Clear Linux.

Also, find here a couple of runs using the i9-13900K box above, but with the always split and never split settings:

• Intel i9-13900K. Always Split.

• Intel i9-13900K. Never Split.

Reproducing the benchmarks is straightforward. First, download the data; the downloaded files will be in the new era5_pds/ directory. Then perform the series of benchmarks; this is takes time, so grab coffee and wait 30 min (fast workstations) to 6 hours (slow laptops). Finally, run the plotting Jupyter notebook to explore your results. If you wish to share your results with the Blosc development team, we will appreciate hearing from you!

Conclusion

Bytedelta can achieve higher compression ratios in most datasets, specially in combination with capable codecs like ZSTD, with a maximum gain of 37% (pressure) over other codecs; only in one case (precip) compression ratio decreases. By compressing data more efficiently, bytedelta can reduce file sizes even more, accelerating transfer and storage.

On the other hand, while bytedelta excels at achieving high compression ratios, this requires more computing power. We have found that for striking a good balance between high compression and fast compression/decompression, other filters, particularly shuffle, are superior overall.

We've learned that no single codec/filter combination is best for all datasets:

• ZSTD (clevel 9) + bytedelta can get better absolute compression ratio for most of the datasets.

• LZ4 + shuffle is well-balanced for all metrics (compression ratio, speed, decompression speed).

• LZ4 (clevel 6) and ZSTD (clevel 1) + shuffle strike a good balance of compression ratio and speed.

• LZ4HC (clevel 6 and 9) + shuffle balances well compression ratio and decompression speed.

• BLOSCLZ without filters achieves best decompression speed (at least in one instance).

In summary, the optimal choice depends on your priorities.

As a final note, the Blosc development team is working on BTune, a new deep learning tuner for Blosc2. BTune can be trained to automatically recognize different kinds of datasets and choose the optimal codec and filters to achieve the best balance, based on the user's needs. This would create a much more intelligent compressor that can adapt itself to your data faster, without requiring time-consuming manual tuning. If interested, contact us; we are looking for beta testers!