High-Performance Data Analysis with the Helmholtz Analytics Toolkit

Kurzfassung

This work introduces the Helmholtz Analytics Toolkit (HeAT), a scientific big data analytics library for High-Performance Computing (HPC) and High-Performance Data Analytics (HPDA) systems. The large progress in big data analytics in general and machine learning/deep learning (ML/DL) in particular, has been considerably enforced by well-designed open source libraries like Hadoop, Spark, Storm, Disco, scikit-learn, H2O.ai, Mahout, TensorFlow, PaddlePaddle, PyTorch, Caffe, Keras, MXNet, CNTK, BigDL, Theano, Neon, Chainer, DyNet, Dask and Intel DAAL to mention only a few of them. Despite the large number of existing data analytics frameworks, a library taking the specific needs in scientific big data analytics under consideration is still missing. For instance, no pre-existing library operates on heterogeneous hardware like GPU/CPU systems while allowing transparent computation on distributed systems. Typical big data analytics frameworks like Spark are designed for distributed memory systems and consequently do not fully explore the shared memory architecture as well as the network technology of HPC systems. ML/DL frameworks like Theano or Chainer focus on single node computations or when providing mechanisms for distributed computation, as done by TensorFlow or PyTorch, they impose the details of the distributed computation to the programmer. Libraries designed for HPC like Dask and Intel DAAL do not provide any GPU support. The presented library - HeAT - is designed for the specific needs of big data analytics in the scientific context. It is based on a distributed tensor data object on which operations can be performed like basic scalar functions, linear algebra algorithms, slicing or broadcasting operations necessary for most data analytics algorithms. The tensor data objects reside either on the CPU or on the GPU and, if needed, are distributed over various nodes. Operations on tensor objects are transparent to the user, i.e. they remain the same irrespective of whether the tensor object resides on a single node or if it is distributed over several nodes allowing to conveniently port algorithms from single nodes to multiple nodes or from CPUs to GPUs. HeAT's tensor module offers a Python-based API almost identical to NumPy, which allow a fast transition from vectorized NumPy code to a parallel and distributed HeAT code. HeAT builds on top of PyTorch, which already provides many required features like automatic differentiation, CPU and GPU support, linear algebra operations, and basic MPI functionality as well as an imperative programming paradigm allowing for fast prototyping essentially in scientific research. In addition to basic tensor operations, HeAT implements several common data analytics algorithms, e.g. k-means, logistic regression and neural networks, optimized for large scale distributed systems. We demonstrate the runtime performance of HeAT using a clustering of 8 GB image data of a rocket engine combustion taken from high-speed cameras. Compared to the performance of the k-means clustering algorithm on MATLAB or the serial HeAT implementation, the distributed computation with 16 MPI ranks leads to a runtime acceleration of about a factor of 20. Since the serial k-means clustering also has a very high memory requirement for these large data sets, the distributed computation even enables the computation of even larger data sets, which would not be possible with single node shared-memory only computation.