January 10, 2020

199 words 1 min read

Transparent Huge Pages on Steroids

Transparent Huge Pages on Steroids

In big memory machines, hugepages can play a large role in increasing system performance. However, using hugepages manually for different segments of memory adds to application complexity. Linux has a …
Talk Title Transparent Huge Pages on Steroids
Speakers Nitin Gupta (Principal Software Engineer, Oracle)
Conference Open Source Summit North America
Conf Tag
Location Los Angeles, CA, United States
Date Sep 10-14, 2017
URL Talk Page
Slides Talk Slides
Video

In big memory machines, hugepages can play a large role in increasing system performance. However, using hugepages manually for different segments of memory adds to application complexity. Linux has a mechanism for automatically backing some memory areas with hugepages, called Transparent Huge Pages (THP). In the current state, THP is quite limited and can only collapse normal pages with a hugepage of one particular size, which is the “default” hugepage size for the system. Such a design is quite limiting for architectures which supported a wide variety of page sizes, ranging from 64KB, all the way to 1TB. In this presentation, Nitin Gupta will discuss ideas for extending THP to support many different page sizes for architectures that support them, along with some performance numbers from initial prototype work.

performance complexity
comments powered by Disqus
Deep reinforcement learning in the enterprise: Bridging the gap from games to industry

Deep reinforcement learning in the enterprise: Bridging the gap from games to industry

December 26, 2019

Mark Hammond explores how enterprises can move beyond games and leverage deep reinforcement learning and simulation-based training to build programmable, adaptive, and trusted AI models for their real-world applications.
Model interpretation guidelines for the enterprise: Using Jupyters interactiveness to build better predictive models (sponsored by DataScience.com)

Model interpretation guidelines for the enterprise: Using Jupyters interactiveness to build better predictive models (sponsored by DataScience.com)

December 21, 2019

Pramit Choudhary offers an overview of Datascience.com's model interpretation library Skater, explains how to use it to evaluate models using the Jupyter environment, and shares how it could help analysts, data scientists, and statisticians better understand their model behaviorwithout compromising on the choice of algorithm.
14 ways to bounce a ball: A tour of the web animations ecosystem

14 ways to bounce a ball: A tour of the web animations ecosystem

December 12, 2019

There are more options than ever for animating things on the web. Bryan Braun leads a tour of the web animations ecosystem, comparing animation approaches in detail and exploring the principles that transcend them all. To do this, Bryan uses lessons learned from building and maintaining Bouncy Ball, a project that compares web animation approaches (think, a TodoMVC for web animation).
Atomic design as a migration strategy

Atomic design as a migration strategy

December 11, 2019

Atomic design is well suited for migrating web applications. Because you build complexity out of simple components, you can start small and slowly carve out your application. Harrison Harnisch shares how Buffer is migrating six years of development with atomic design.
Big data computations: Comparing Apache HAWQ, Druid, and GPU databases

Big data computations: Comparing Apache HAWQ, Druid, and GPU databases

December 5, 2019

The class of big data computations known as distributed merge trees was built to aggregate user information across multiple data sources in the media domain. Vijay Srinivas Agneeswaran explores prototypes built on top of Apache HAWQ, Druid, and Kinetica, one of the open source GPU databases. Results show that Kinetica on a single G2.8x node outperformed clusters of HAWQ and Druid nodes.
Hadoop and object stores: Can we do it better?

Hadoop and object stores: Can we do it better?

December 3, 2019

Trent Gray-Donald and Gil Vernik explain the challenges of current Hadoop and Apache Spark integration with object stores and discuss Stocator, an open source object store connector that overcomes these shortcomings by leveraging object store semantics. Compared to native Hadoop connectors, Stocator provides close to a 100% speedup for DFSIO on Hadoop and a 500% speedup for Terasort on Spark.