Btune: Making Compression Better

What It Is?​

Btune is a plugin for Blosc2 that can help you finding the optimal combination of compression parameters for your datasets. Depending on your needs, Btune has three different tiers of support, each with its own advantages and use cases:

• Genetic (Btune Community): This genetic algorithm tests different combinations of compression parameters to meet the user's requirements for both compression ratio and speed for each chunk in the dataset. It assigns a score to each combination and, after a number of iterations, the software stops and uses the best score (minimal value) found for the rest of the dataset. For a graphical visualization, click on the image, select an example, and click on the 'play' button (it may require clicking twice). This is best suited for personal use.

• Trained (Btune Models): With this approach, the user sends a representative sample of datasets to the ironArray team and receives back trained neural network models that enable Btune to predict the best compression parameters for similar or related datasets. This approach is best for workgroups that need to optimize for a limited variety of datasets.

• Fully managed (Btune Studio): The user receives a license to use our training software, which enables on-site training for an unlimited number of datasets. The license also includes a specified number of training/consultancy hours to help the user get the most out of the training process. Refer to the details below for more information. This approach is best suited for organizations that need to optimize for a wide range of datasets.

Why Btune?​

Essentially, because compression is not a one-codec-fits-all problem. Compressing data involves a trade-off between compression ratio and speed. A higher compression ratio results in a slower compression process. Depending on your needs, you may want to prioritize one over the other.

Finding the optimal compression parameters in Blosc2 can be a slow process due to the large number of combinations of compression parameters (codec, compression level, filter, split mode, number of threads, etc.), and it may require a significant amount of manual trial and error to find the best combinations. However, you can significantly speed up this process by using Btune while compressing your datasets.

For instance, if you are storing data from high-speed data acquisition systems, you may want to prioritize compression speed over compression ratio. This is because you will be writing data at speeds near the capacity of your systems. On the other hand, if the goal is to access the data repeatedly from a file system, you may want to prioritize decompression speed over compression ratio for optimal performance.

See for instance the following figures that show the trade-offs between compression ratio and speed for different codecs and filters (these are for chunks of weather data):

And here, the different codecs and filters are compared in terms of compression ratio:

With Btune, you can find the optimal combination of compression parameters (aka in the Pareto front) for your datasets, allowing you to achieve the best possible compression ratio and speed for your specific needs.

How To Use​

Btune is a plugin for Blosc2 that can be obtained from the PyPI repository. You can learn how to use it in the Btune README. The plugin is currently only available for Linux and macOS, and only for Intel architecture. However, we plan to add support for other architectures in the future.

The Btune plugin above can be used for both Btune Community and Btune Models. For Btune Studio, you will need to contact us to get the additional software for training the models.

Here you have a couple of tutorials covering different aspects of Btune:

• Btune Genetic & Trained
• Btune Studio

For completing the Studio version, you will need to contact us to get the additional software for training the models.

What's in a Model?​

A neural network is a simplified model of the way the human brain processes information. It simulates a large number of interconnected processing units that resemble abstract versions of neurons. These processing units are arranged in layers, which are connected by weights that are adjusted during the training process. To train the network, a large number of examples are fed into it, and the weights are adjusted to minimize the difference between the expected output and the actual output. Once training is complete, the network can be used to predict the output for new inputs.

In our context, the "model" refers to the serialization of the layers and weights of the trained neural network. It is delivered to you as a set of small files (in JSON and TensorFlow format) that can be placed anywhere in your filesystem for Btune to access. By using this model, Btune can predict the optimal combination of compression parameters for a given chunk of data. The inference process is very fast, making it suitable for selecting the appropriate compression parameters on a chunk-by-chunk basis while consolidating large amounts of data.

A Starry Example​

In the figure below, you can see the most predicted combinations of codecs and filters when optimizing for decompression performance on a subset of the Gaia dataset. The subset contains stars that are less than 10,000 light years away from our Sun (around 500 millions). The data is stored in an array of shape (20,000, 20,000, 20,000), with the number of stars in every cubic light year cell, resulting in a total uncompressed size of 7.3 TB.

Now, the following figure displays the speed that can be achieved by obtaining multiple multidimensional slices of the dataset along different axes, using the most efficient codecs and filters for various tradeoffs. The speed is measured in GB/s, so a higher value is better.

The results indicate that the fastest compression combination is BloscLZ (compression level 5), closely followed by Zstd (compression level 9). Also, note how the fastest codecs, BloscLZ and also Zstd, are not affected very much by the number of threads used, which means that they are not CPU-bound, so small computers or laptops with low core counts will be able to reach good speeds.

Finally, it is important to compare the compression ratios achieved by different codecs and filters. In the following figure, we can see the file sizes created when using the most commonly predicted codecs and filters for various trade-offs. The file sizes are measured in GB, so the lower, the better.

In this case, the trained model recommends using Zstd (compression level 9) for a good balance between compression ratio and decompression speed, and that can be confirmed by seeing the large difference in size. However, note that BitShuffle + Zstd (compression level 9) is not a good option in general, unless you are looking for the absolute best compression ratio.

You can read more context about this example in our paper for SciPy 2023 (talk slides here). The data and scripts used are available here.

Licensing Model​

There are different licenses available for Btune.

Btune Community allows you to explore compression parameters that are better suited to your datasets. However, this process can be slow and may require a large number of iterations before finding the best combination. Additionally, certain chunks in the same dataset may benefit more from a particular combination, while others may benefit more from a different one.

Btune Models addresses the limitations of Btune Community by automatically finding the best combination for chunks in a dataset, without requiring any manual operation. This is made possible by using pre-trained neural network models, which allow the best combination to be found on a chunk-by-chunk basis, thereby increasing the effectiveness of the compression process.

Btune Studio goes one step further and provides access to the necessary software for training the datasets by yourself. In this way, you have control over all the necessary components to find optimal compression parameters on your own (but you can still buy additional support from us, if needed).

Pricing​

Btune Community​

Free to use (but hey, if you like the project, please consider donating too). Note that Btune is licensed under an Affero GPLv3 license. This license comes with limited support, as it is mostly a community-supported project.

Btune Models​

Requires a fee of $3500 USD (or 3500 EUR) for up to 4 trained models. If you need more, ask for a quote.

Btune Studio​

Requires a fee of $9500 USD (or 9500 EUR) for the first year, or $950 USD (or 950 EUR) per month for at least 1 year, whichever fits best for you. Renewal is $7500 USD (or 7500 EUR) per year, or $750 USD (or 750 EUR) monthly after the 1st year. If you don't renew, you keep the right to use Btune Studio for producing models internally in your organization forever; however, you won't have access to newer versions.

Note: With Btune Studio we deliver sources of it, so that you can build/fix it yourself. However, you won't be able to include it in your own software and distribute it without permission.

Priority Support​

We offer different packages for support. The support can be used for training in the use of the software, or for consultation on compression for big data in general.

Donations​

If you prefer donating instead (e.g. for the Community Edition), you can do that via the donations form for the Blosc project.

Testimonials​

Blosc2 and Btune are fantastic tools that allow us to efficiently compress and load large volumes of data for the development of AI algorithms for clinical applications. In particular, the new NDarray structure became immensely useful when dealing with large spectral video sequences.

-- Leonardo Ayala, Div. Intelligent Medical Systems, German Cancer Research Center (DKFZ)

Btune is a simple and highly effective tool. We tried this out with @LEAPSinitiative data and found some super useful spots in the parameter space of Blosc2 compression arguments! Awesome work, @Blosc and @ironArray teams!

-- Peter Steinbach, Helmholtz AI Consultants Team Lead for Matter Research @HZDR_Dresden

Contact​

If you are interested in Btune and have any further questions, please contact us at contact@ironarray.io.