drj logo
  • This field is for validation purposes and should be left unchanged.

Already have an account? Log in

drj logo

Welcome to DRJ

Already registered user? Please login here

Login Form

Register
Forgot password? Click here to reset

Create new account
(it's completely free). Subscribe

Skip to content
Disaster Recovery Journal
  • EN ESPAÑOL
  • SIGN IN
  • SUBSCRIBE
  • THE JOURNAL
    • Young Professional Spotlight
    • Career Spotlight
    • Article Submission
    • Digital Edition
    • DRJ Annual Resource Directories
    • Article Archives
  • EVENTS
    • Spring 2021
    • Call for Papers – Fall 2021
    • DRJ Town Hall Meetings
    • Other Industry Events
    • Schedule & Archive
  • WEBINARS
    • Upcoming
    • On Demand
  • MENTOR PROGRAM
  • DRJ ACADEMY
  • RESOURCES
    • White Papers
    • DR Rules and Regs
    • Industry Groups
    • Business Directory
    • Business Resilience Decoded
    • DRJ Glossary of Business Continuity Terms
  • ABOUT
    • Board and Committees
      • Executive Council Members
      • Editorial Advisory Board
      • Career Development Committee
      • FAQs Committee
      • Web Committee
      • Glossary Committee
      • Rules and Regulations Committee
  • Podcast

Tech Alert: Reducing Operational Costs of AI/ML Architectures

by Jon Seals | July 23, 2019 | Uncategorized | 0 comments

SANTA CLARA, Calif. – Artificial Intelligence and Machine Learning applications are becoming ubiquitous in companies of all sizes, across all industries, and require significant up-front investment in hardware. Along with initial costs, however, come ongoing operational costs of running these data-intensive workloads. Experts at Quobyte® Inc., a leading developer of modern storage system software, offer the following tips for reducing the operational expenses of AI/ML infrastructures.

Smart storage

AI/ML workloads have different performance profiles, including the high-throughput, low-latency requirements in the model training stage, plus there can be large-block sequential, small-block random, or mixed general workloads during the ingest or other stages. The storage must be performant enough to keep up with the varying requirements. GPUs are one of the most expensive assets in the whole system and should be kept busy, not wastefully underutilized.

Avoid point solutions

There should be no need to build and maintain a separate storage architecture for AI/ML. It’s tempting to keep production workloads on a different system, but cobbling together two or three products to handle AI/ML data requirements adds a layer of complexity, and incurs management and maintenance costs. Instead of an isolated infrastructure, use the performance and capacity resources you have.

No more tiers

Likewise, avoid cumbersome tiering. A unique characteristic of AI/ML workloads is that all data is conceivably hot data. Training data is used very frequently, and should remain accessible. Typical tiering strategies introduce bottlenecks and can ultimately add to the overall costs, when you probably intended to save. Rather than using different storage based on whether data is hot or cold, mix HDD and SSD to get the best price/performance for your needs.

Eliminate migration

With one tier-free, unified storage infrastructure throughout the data center, you can eliminate the time-consuming and error-prone process of copying data from stage to stage and locality to locality. Data should never leave the file system at any point during the entire AI/ML lifecycle.

Consider cost of scale

AI/ML data sets easily grow to the hundreds of petabytes. A high-capacity storage solution may be as easy to manage as a single box today, but will require exponentially more manual labor when you have 200 of them. Maintenance looks very different at scale when drive failures, network issues, broken hardware, or updates become a daily occurrence.

Be ready for change

As project requirements change – sometimes more quickly than anticipated – the infrastructure must adapt. Downtime equals dollars wasted as GPUs sit idle. You should be able to add disks or servers when you need more performance, without any interruption to applications or services. Storage software can automatically detect and remove broken hardware; self-healing capabilities will use available resources elsewhere in the cluster until it can be replaced when convenient.

“Hardware and GPUs are a big up-front investment, but there are running costs to consider and the way the AI/ML infrastructure is built can add to or mitigate it,” said Bjorn Kolbeck, Quobyte CEO. “Containing ongoing costs starts with getting the storage performance to take full advantage of your investment, and building a flexible environment that grows when needed without a maintenance and management overhead.”

Follow Quobyte

https://www.twitter.com/quobyte
https://www.linkedin.com/company/quobyte
https://www.facebook.com/quobyte

About Quobyte

Building on a decade of research and experience with the open-source distributed file system XtreemFS and from working on Google’s infrastructure, Quobyte delivers on the promise of software-defined storage for the world’s most demanding application environments including High Performance Computing (HPC), Machine Learning (ML), Media & Entertainment (M&E), Life Sciences, Financial Services, and Electronic Design Automation (EDA). Quobyte uniquely leverages hyperscaler parallel distributed file system technologies to unify file, block, and object storage. This allows customers to easily replace storage silos with a single, scalable storage system – significantly saving manpower, money, and time spent on storage management. Quobyte allows companies to scale storage capacity and performance linearly on commodity hardware while eliminating the need to expand administrative staff through the software’s ability to self-monitor, self-maintain, and self-heal. Please visit www.quobyte.com for more information.

Recent Posts

  • Download Your Planning Templates
  • They Called Me ‘The Queen of Doom’
  • Business Continuity as a Business Strategy
  • AlertMedia Improves Emergency Management with the Launch of Event Pages: Connecting People with Critical Information in Real-Time
  • Disaster Recovery Planning in a VM Operating Environment

Recent Comments

    Archives

    • December 2020 (1)
    • October 2020 (1)
    • August 2020 (1)
    • January 2020 (1)
    • November 2019 (9)
    • October 2019 (8)
    • September 2019 (8)
    • August 2019 (17)
    • July 2019 (77)
    • June 2019 (92)
    • May 2019 (128)
    • April 2019 (109)
    • March 2019 (80)
    • February 2019 (106)
    • January 2019 (102)
    • December 2018 (96)
    • November 2018 (131)
    • October 2018 (139)
    • September 2018 (81)
    • August 2018 (124)
    • July 2018 (113)
    • June 2018 (90)
    • May 2018 (142)
    • April 2018 (151)
    • March 2018 (124)
    • February 2018 (157)
    • January 2018 (140)
    • December 2017 (118)
    • November 2017 (158)
    • October 2017 (186)
    • September 2017 (139)
    • August 2017 (212)
    • July 2017 (179)
    • June 2017 (214)
    • May 2017 (230)
    • April 2017 (223)
    • March 2017 (159)
    • February 2017 (175)
    • January 2017 (112)
    • December 2016 (8)
    • October 2016 (12)
    • June 2016 (15)
    • May 2016 (4)
    • April 2016 (10)
    • January 2016 (10)
    • December 2015 (2)
    • October 2015 (12)
    • September 2015 (2)
    • June 2015 (15)
    • April 2015 (15)
    • March 2015 (6)
    • February 2015 (7)
    • January 2015 (6)
    • December 2014 (18)
    • November 2014 (19)
    • October 2014 (24)
    • September 2014 (11)
    • August 2014 (13)
    • July 2014 (7)
    • June 2014 (36)
    • May 2014 (19)
    • April 2014 (21)
    • March 2014 (90)
    • February 2014 (5)
    • January 2014 (26)
    • December 2013 (10)
    • November 2013 (25)
    • October 2013 (37)
    • September 2013 (27)
    • August 2013 (8)
    • July 2013 (8)
    • June 2013 (31)
    • May 2013 (17)
    • April 2013 (103)
    • March 2013 (82)
    • February 2013 (120)
    • January 2013 (212)
    • December 2012 (128)
    • November 2012 (92)
    • October 2012 (159)
    • September 2012 (112)
    • August 2012 (153)
    • July 2012 (106)
    • June 2012 (91)
    • May 2012 (35)
    • April 2012 (49)
    • March 2012 (14)
    • February 2012 (19)
    • January 2012 (44)
    • December 2011 (9)
    • November 2011 (21)
    • October 2011 (45)
    • September 2011 (13)
    • August 2011 (9)
    • July 2011 (47)
    • May 2011 (46)
    • April 2011 (10)
    • March 2011 (25)
    • February 2011 (1)
    • January 2011 (17)
    • December 2010 (8)
    • October 2010 (19)
    • September 2010 (1)
    • July 2010 (23)
    • June 2010 (1)
    • May 2010 (1)
    • April 2010 (21)
    • March 2010 (1)
    • February 2010 (1)
    • January 2010 (4)
    • December 2009 (20)
    • November 2009 (4)
    • October 2009 (17)
    • September 2009 (11)
    • August 2009 (2)
    • July 2009 (2)
    • June 2009 (79)
    • April 2009 (20)
    • March 2009 (1)
    • January 2009 (22)
    • November 2008 (1)
    • October 2008 (1)
    • July 2008 (4)
    • June 2008 (22)
    • April 2008 (25)
    • February 2008 (1)
    • January 2008 (16)
    • December 2007 (17)
    • November 2007 (662)
    • October 2007 (410)
    • September 2007 (1)
    • November 1999 (1)

    ARTICLES & NEWS

    • Business Continuity
    • Disaster Recovery
    • Crisis Management & Communications
    • Risk Management
    • Article Archives
    • Industry News

    THE JOURNAL

    • Digital Edition
    • Young Professionals
    • Career Spotlight
    • Advertising & Media Kit
    • Submit an Article

    RESOURCES

    • White Papers
    • Rules & Regulations
    • FAQs
    • Glossary of Terms
    • Industry Groups
    • Business & Resource Directory
    • Business Resilience Decoded

    EVENTS

    • Spring 2020
    • Fall 2020

    WEBINARS

    • Watch Now
    • Upcoming

    CONTACT

    • Article Submission
    • Connect with our Audience
    • Contact Us

    ABOUT DRJ

    Disaster Recovery Journal is the industry’s largest resource for business continuity, disaster recovery, crisis management, and risk management, reaching a global network of more than 138,000 professionals. Offering weekly webinars, the latest industry news, rules and regulations, podcasts, the industry’s only official mentoring program, a quarterly magazine, and two annual live conferences, DRJ is leading the way to keep professionals up-to-date and connected in an ever-changing world.

    LEARN MORE

    TWITTER

    Disaster Recovery Journal is the leading publication/event covering business continuity/disaster recovery.

    Follow us for daily updates @drjournal

    Newsletter

    The Journal, right in your inbox.

    Be informed and stay connected by getting the latest in news, events, webinars and whitepapers on Business Continuity and Disaster Recovery.

    Subscribe Now
    Copyright 2021 Disater Recovery Journal
    • Terms of Use
    • Privacy Policy