Welcome to The Long View—where we peruse the news of the week and strip it to the essentials. Let’s work out what really matters.
This week: New ideas bring low-power ML inference, and more big-tech jobs are going. (more…)
Tag: deep learning
-
8-Bit Floating Point for AI/ML? | Amazon and Microsoft Shed Tech Jobs
-
Databricks Unified Analytics Platform Simplifies Distributed Deep Learning
Unified Analytics Leader Supports New Apache Spark 2.4; Introduces New Feature to Simplify Distributed Deep Learning
SAN FRANCISCO–(BUSINESS WIRE)–Databricks, the leader in unified analytics and founded by the original creators of Apache Spark™, today announced support for the newly released Apache Spark 2.4.0 within Databricks’ Unified Analytics Platform. Databricks is the first unified analytics vendor to support Apache Spark 2.4. It is supported as part of Databricks Runtime 5.0, which is now generally available. Databricks also introduced a key feature, HorovodRunner, within Runtime 5.0 to further simplify distributed deep learning.
The Apache Spark community made multiple valuable contributions to the Spark 2.4 release which was introduced on November 8, 2018. In this release, Project Hydrogen substantially improves the performance and fault-recovery of distributed deep learning and machine learning frameworks on Spark. Project Hydrogen directly addresses the challenges data teams face because there is a significant difference between how big data jobs and deep learning jobs are executed. Whereas Spark excels at data processing at massive scale, deep learning assumes complete coordination and dependency among tasks which is optimized for constant communication rather than scalability and fault tolerance.
“Innovation continues to thrive within the Apache Spark community. Project Hydrogen is the most recent major initiative with an aim to provide first-class support for popular distributed machine learning frameworks on Apache Spark,” said Reynold Xin, co-founder at Databricks, Apache Spark PMC member and the top contributor to the project.
Within Apache Spark 2.4, Project Hydrogen introduces Barrier Execution, a new scheduling mode that allows practitioners to properly embed distributed deep learning training as an Apache Spark workload. Added Xin, “This is the largest change to Spark’s scheduler since the inception of the project. At Databricks, we also found additional opportunities to simplify the complexity of machine learning workloads. Within Databricks’ Unified Analytics Platform, which is powered by Spark 2.4, we created further optimizations to simplify distributed deep learning.”
 Databricks Simplifies Distributed Deep Learning in Runtime 5.0
Model experimentation usually takes place on a single-node machine, locally or in the cloud, before scaling out computation as needed. Migrating from single-node workloads to distributed training on a CPU or GPU clusters can often times require a full code rewrite, increasing the complexity of moving to distributed training. To accelerate migration to distributed deep learning, Databricks just released HorovodRunner. The new feature provides a simple way to scale up deep learning training workloads from a single machine to large clusters, reducing overall programming and training time from hours to minutes.
To help simplify deep learning further, Databricks also provides native integration with the most popular frameworks including TensorFlow, Keras, and Horovod, as well as a performance edge with the most popular machine learning algorithms from MLlib and GraphFrames. This provides practitioners with a convenient way to get machine learning clusters started in seconds, pre-configured with the latest machine learning frameworks, libraries, and their dependencies.
Additional Resources
- Databricks Blog:Â Announcing Databricks Runtime 5.0
- Databricks Blog:Â Introducing Apache Spark 2.4
About Databricks
Databricks’ mission is to accelerate innovation for its customers by unifying Data Science, Engineering and Business. Founded by the original creators of Apache Spark, Databricks provides a Unified Analytics Platform for data science teams to collaborate with data engineering and lines of business to build data products. Users achieve faster time-to-value with Databricks by creating analytic workflows that go from ETL and interactive exploration to production. The company also makes it easier for its users to focus on their data by providing a fully managed, scalable, and secure cloud infrastructure that reduces operational complexity and total cost of ownership. Databricks, venture-backed by Andreessen Horowitz, NEA and Battery Ventures, among others, has a global customer base that includes Viacom, Shell and HP.
Apache, Apache Spark and Spark are trademarks of the Apache Software Foundation.
-
How AI Impacts DevOps Pipelines
The rise of machine- and deep-learning algorithms is about to create some interesting data-gravity challenges for IT operations teams. Over most of the last decade IT operations teams have been caught up in a public- versus private-cloud computing debate that has mostly been species. The truth is that most organizations wind up employing both. But as IT organizations are just now getting used to that multi-cloud reality, the rise of machine- and deep-learning algorithms to create a new generation of advanced artificial intelligence (AI) applications are about to add some significant nuances to managing the DevOps pipeline.
New applications tend to be built and deployed where they can access the most relevant data. Most of the data employed by the average enterprise resides on-premises today. But it’s clear that the data being created in various types of cloud services is growing at a faster rate, thereby altering data gravity forces across the enterprise.
To get the most value out of that, developers are now building a new generation of advanced applications that employ machine- and deep-learning algorithms. In most cases those algorithms are being deployed wherever data happens to reside. For example, Salesforce just announced at its Salesforce TRAILHEADX 2017 conference that it is making available additional APIs to the Salesforce Einstein Platform Services for building custom applications that can now be infused with sentiment analysis, the ability to detect objects and understand customer intent.
Michael Machado, director of product management for Salesforce Einstein, says that the success of any AI initiative will come down to both the amount and quality of the data being accessed. As a CRM application consisting of a mix of structured and unstructured data concerning customers, Salesforce represents an attractive platform for building AI applications. In fact, to facilitate the building of applications, Salesforce also is making it simpler for developers to collaborate across multiple alliances in addition to providing access to DevOps tools from Atlassian and Github.
But CRM applications are only one element of the typical application portfolio. The more data that any set of algorithms can be applied against, the faster the application learns and the more accurate it becomes. As developers build the next generation of applications, they are going to want to be able to access as much data as possible. For many IT organizations, that will mean the applications will be hybrid almost by definition—attempting to centralize all their data will prove to be impractical as well as prohibitively expensive. Constructing application pipelines spanning multiple data sources, however, requires a very mature set of DevOps processes.
Arguably, the rise of these next-generation applications might finally force the DevOps issue inside most organizations. Most have been able to manage application development on various cloud platforms in isolation. Going forward, that may no longer be practical, as developers seek to incorporate more data sources within their application environment. The challenge IT operations teams will need to wrestle with is providing timely access to that data in way that doesn’t ultimately compromise their performance.
— Mike Vizard
-
The Time Is Now for an IT Operations Revolution
There’s a level of frustration starting to build inside IT operations teams that’s about to boil over. Lucas Carlson, vice president of strategy for Automic, a unit of CA Technologies that provides an IT automation platform, says developers are bullying IT operations teams by programmatically invoking application programming interfaces (APIs) exposed by public cloud service providers. As the number of application workloads running in those clouds increases, IT operations professionals are losing control.
To ensure their continued relevance, IT operations teams need to leverage automation to inject levels of agility into local data center environments, he says. That would make it much less likely developers would feel the need to programmatically invoke IT infrastructure resources.
Carlson says achieving that level of agility will require IT operations teams to rely more on machine- and deep-learning algorithms to automate low-level tasks. Instead of thinking in terms of managing silos of compute, storage and networking resources, Carlson contends IT operations professionals need to reinvent themselves.
The artificial intelligence (AI) models based on advanced algorithms must be trained continuously. IT operation professionals could become the trainers of those AI models. In fact, Carlson says IT operations teams are approaching a critical moment in history: They can either wait for AI models to eventually replace them or they can leverage them to take back control of IT infrastructure.
However, developers tend to make decisions based on what they prefer versus what’s best for the organization. Before too long, multiple duplicate operations management frameworks are in place. Rather than incur all those management headaches and associated technical debt, Carlson says organizations need a single framework that can more adroitly accommodate the multiple roles within the DevOps team. It’s already long past the time when IT teams should have modernized their operations, he says.
The relationship between developers and IT operations teams has always been contentious. The more belligerent members of the developer community are even making a case for “NoOps.” In that context, all the IT dollars spent on operations are moved into developing applications because the management of IT infrastructure is completely automated. Given the amount of legacy infrastructure that isn’t likely changing anytime soon, achieving that level of automation inside the enterprise won’t happen for a while. But there’s no doubt that IT operations needs to get more agile. Because of that issue, investment in software-defined data center (SDDC) technologies is growing rapidly.
Instead of resisting that trend, Carlson says teams need to demand that internal IT operations be brought into the 21st century. It’s true that SDDCs infused with AI models will automate most of the lower-level tasks associated with DevOps. But those members of that IT operations teams who become the masters of building and training those AI models will be among the most highly sought-after individuals in IT.
— Mike Vizard