Introducing the Intel® Data Analytics Acceleration Library

The Intel® Data Analytics Acceleration Library (Intel® DAAL) is the library of Intel® Architecture optimized building blocks covering all data analytics stages: data acquisition from a data source, preprocessing, transformation, data mining, modeling, validation, and decision making. To achieve best performance on a range of Intel® processors, Intel DAAL uses optimized algorithms from the Intel® Math Kernel Library and Intel® Integrated Performance Primitives.

Intel DAAL supports the concept of the end-to-end analytics when some of data analytics stages are performed on the edge devices (close to where the data is generated and where it is finally consumed). Specifically, Intel DAAL Application Programming Interfaces (APIs) are agnostic about a particular cross-device communication technology and therefore can be used within different end-to-end analytics frameworks.

Intel DAAL consists of the following major components: Data Management, Algorithms, and Services:

The Data Management component includes classes and utilities for data acquisition, initial preprocessing and normalization, for data conversion into numeric formats done by one of supported Data Sources, and for model representation. The NumericTable class and its derivative classes within the Data Management component are intended for in-memory numeric data manipulation. The Model class mimics the actual data and represents it in a compact way so that you can use the library when the actual data is missing, incomplete, noisy, or unavailable. These are the key interfaces for processing a data set with algorithms. The DataSourceDictionary and NumericTableDictionary classes provide generic methods for dictionary manipulation, such as accessing a particular data feature, setting and retrieving the number of features, and adding a new feature.

The Algorithms component consists of classes that implement algorithms for data analysis (data mining), and data modeling (training and prediction). These algorithms include matrix decompositions, clustering, classification, and regression algorithms, as well as association rules discovery.

Algorithms support the following computation modes:

• Batch processing
• Online processing
• Distributed processing

In the batch processing mode, the algorithm works with the entire data set to produce the final result. A more complex scenario occurs when the entire data set is not available at the moment or the data set does not fit into the device memory.

In the online processing mode, the algorithm processes a data set in blocks streamed into device memory by doing incrementally updating partial results, which are finalized upon processing of the last data block.

In the distributed processing mode, the algorithm operates on a data set distributed across several devices (compute nodes). The algorithm produces partial results on each node, which are finally merged into the final result on the master node.

Depending on the usage, algorithms operate both on actual data (data set) and data models. Analysis algorithms typically operate on data sets. Training algorithms typically operate on a data set to train the appropriate data model. Prediction algorithms typically work with the trained data model and with a working data set.

Models produced by training algorithms are associated with interfaces for generation of metrics that characterize the models. These metrics can be used during creation of a model to understand its quality, or in a production mode to track agreement between actual data and the model that represents it.

The Services component includes classes and utilities used across Data Management and Algorithms components. These classes enable memory allocation, error handling, and implementation of collections and shared pointers. A collection enables storing different type of objects in a unified manner. In Intel DAAL, collections are used for handling the input and output of algorithms and for error handling. Intel DAAL implements shared pointers to enable memory handling needed for memory management operations such as deallocation.

Classes implemented in Data Management, Algorithms, and Services components cover most important usage scenarios and allow seamless implementation of complex data analytics workflows through direct API calls. At the same time the library is an object-oriented framework that enables you to customize the API by redefining particular classes and methods of the library.