Your Browser is Out of Date

Nytro.ai uses technology that works best in other browsers.
For a full experience use one of the browsers below
Accelerating End-to-End Machine Learning Using the Intel® oneAPI Toolkit and Intel® Flex Series GPU
Executive summary Introduction
Introduction
Evaluation of oneAPI compatible ETL libraries Optimized End-to-End ML Pipeline vs. Serial Pipeline Evaluation of ML Pipeline with Varying Dataset Sizes Conclusions Appendix: Reproducibility References

Introduction

Introduction

  • An end-to-end (E2E) Machine Learning (ML) pipeline consists of three general stages: (1) Extraction, Transfer, and Loading (ETL) of data; (2) Model Training; and (3) Inference and Visualization, as illustrated in Figure 1. These stages are generally computationally expensive, which impacts the performance and effectiveness of the resulting ML pipeline.

    The goal of this project is to accelerate the performance of an end-to-end ML pipeline using Intel’s oneAPI AI Analytics Toolkit [1]. The oneAPI AI Analytics toolkit is a set of powerful and familiar Python packages and frameworks which are optimized to deliver drop-in acceleration for Intel architectures. With oneAPI libraries built for low-level compute optimizations, it is possible to achieve highly efficient multithreading, vectorization, and memory management, and to scale scientific computations efficiently across a cluster.

    Figure 1.   End-to-end ML Pipeline using Intel’s oneAPI AI Analytics Toolkit

    Some of the key features of this toolkit, as shown in Figure 1, are:

    1. Data Analytics at Scale:
    • Delivers an excellent out-of-the-box developer experience using open, optimized software tools coupled with optimal compute and memory hardware configurations for AI applications.
    • Considerable reduction in time for preprocessing: sort, filter, label. and transform data using high-memory systems that can accommodate large datasets efficiently.
    • Accelerate pandas workloads across multiple cores and multiple nodes using Intel’s distribution of Modin.
    1. Optimized Frameworks and Middleware:
    • Efficient data transfer performance across packages through efficient copy data exchange with data-parallel Python technology.
    • Intel optimized frameworks for TensorFlow and PyTorch, pre-trained models, and low-precision tools to deliver high-performance training on Intel XPUs and incorporate fast inference into the AI development workflow.
    • Integrated Intel performance libraries such as Intel® oneMKL (Math Kernel Library) and Intel® oneDAL (Data Analytics Library) to accelerate NumPy, scikit-learn and XGBoost.

    The oneAPI AI Analytics Toolkit [1] is implemented using the oneAPI Data Analytics Library (oneDAL), a powerful machine learning library that helps speed up big data analysis. oneDAL is an extension of Intel® Data Analytics Acceleration Library (DAAL) and is a part of oneAPI. oneAPI is a cross-industry, open, standards-based unified programming model that delivers a common developer experience across accelerator architectures. For more details about these libraries, see the References section for oneAPI [2] and oneDAL [3].

    This study was conducted using two datasets: the New York City (NYC) Taxi historical dataset [4] and the Higgs dataset [5]:

    • The NYC dataset is used to study the performance of the regression function for machine learning. The data was collected and provided to the NYC Taxi and Limousine Commission (TLC) by technology providers authorized under the Taxicab & Livery Passenger Enhancement Programs (TPEP/LPEP). It contains historical records accumulated and saved on individual monthly files from 2014 to 2016 (total: ~64GB).
    • The Higgs dataset is used for classification problems to distinguish between a signal process which produces Higgs bosons and a background process which does not. The data was produced using Monte Carlo simulations. The first 21 features (columns 2-22) are kinematic properties measured by the particle detectors in the accelerator. The last seven features are functions of the first 21 features and are high-level features derived by physicists to help discriminate between the two classes. The Higgs dataset has 11 million instances (total: ~7.5GB).

    As stated above and illustrated in Figure 1, the E2E ML pipeline consists of three general stages. The Extraction, Transfer, and Loading (ETL) stage is the first stage of the pipeline. In this stage, the data is extracted from different sources and transformed into the required format to be used for the Model Training stage. Because the ETL stage is the most time-consuming stage in the pipeline, it is essential to optimize this stage. In the section Evaluation of oneAPI compatible ETL libraries in this paper, we describe the experiments that we performed to evaluate the various oneAPI-compatible ETL libraries. We compared the ETL performance when implemented using Modin [6] (with Ray, Dask, and Omnisci) and the native pandas library. The performance comparison is further broken down into the most commonly used ETL functions: df.read_csv, df.rename, df.drop, df.fill_na, and df.concat. Based on the performance advantage of Modin-Dask, we optimized the ETL stage of our ML pipeline using the Modin-Dask library.

    The optimized end-to-end ML pipeline was implemented and deployed on a computing node of the Intel® DevCloud [7], containing an Intel® Ice Lake CPU (with 96 CPU cores) and a 150W Intel® Arctic Sound GPU. In the section Optimized End-to-End ML Pipeline vs. Serial Pipeline, the performance of the optimized implementation (using Modin-Dask and oneDAL) is compared against the unoptimized version (using native panda and XGBoost libraries), with a performance breakdown by the three stages in the pipeline.  The comparison is performed for both a NYC dataset (for ML regression analysis) and a Higgs dataset (for classification problem). In the section Evaluation of ML Pipeline with Varying Dataset Sizes, the optimized code is then used to study the pipeline performance based on varying dataset size. The NYC dataset was varied from 2 GB to 64 GB (increment of power of 2). The Higgs dataset was varied from 1 GB to 7.5 GB (increment of 1 GB).