begriffs

Machine Learning at the Limit

March 13, 2015

Professor John Canny presents benchmarks for the BIDMach Scala machine learning library. He and his team achieve a nearly theoretical maximum speed in this library relative to roofline analysis of the available CPU and GPU resources. They have managed to outperform entire machine learning clusters using a single computer. Check out the video below for more detailed comparisons.

This talk was delivered at Alpine Data Lab.

Slides

Summary

  • Discovering the limits of machine learning performance using roofline design from computer architecture
  • Taking fully accelerated single-node algorithms and scaling them up
  • Tweaking a data model for machine learning should be quick and iterative
  • Often prototype machine learning is done in Scikit-Learn or R. It is then translated for production code by a group that often lacks machine learning expertise, which degrades the quality of results.
  • The reasons for developing a new ML toolkit (Bidmach):
    • GPUS now operate efficiently on sparse data
    • minibatch and stochastic gradient descent habe become a workhorse of choice for big data processing
    • easy customization of models
    • quick exploration, and sharing code between prototype and production
  • To design the framework we start with roofline design and articulating the limits of CPUs etc
  • We build the high level operations on top of a GPU-accelerated matrix library
  • Roofline design establishes fundamental performance limits for a computational kernel
  • Comparison of the strengths of CPUs and GPUs
  • Comparing performance of BIDMach vs Vowpal Wabbit vs Scikit-Learn
  • For many tasks a single node running BIDMach outperforms a whole cluster running traditional libraries
  • Measurements of running time for Latent Dirichlet Allocation, random forests, single-class regression, logistic regression, multilabel regression, factor models, SVMs and clustering
  • Rooflining network bandwidth