Frank McQuillan

In this session we will present an efficient way to train many deep learning model configurations at the same time with Greenplum, a free and open source massively parallel database based on PostgreSQL. The implementation involves distributing data to the workers that have GPUs available and hopping model state between those workers, without sacrificing reproducibility or accuracy. Then we apply optimization algorithms to generate and prune the set of model configurations to try.

Deep neural networks are revolutionizing many machine learning applications, but hundreds of trials may be needed to generate a good model architecture and associated hyperparameters. This is the challenge of model selection. It is time consuming and expensive, especially if you are only training one model at a time.

Massively parallel processing databases can have hundreds of workers, so can you use this parallel compute architecture to address the challenge of model selection for deep nets, in order to make it faster and cheaper?

It’s possible!

We will demonstrate results from this project using a version of Hyperband, which is a well known hyperparameter optimization algorithm, and the deep learning frameworks Keras and TensorFlow, all running on Greenplum database using Apache MADlib. Other topics will include architecture, scalability results and bright opportunities for the future.

We look forward to presenting this topic at FOSDEM’20!

In this session we will discuss the use of massively parallel databases for deep learning, drawing on experience from running deep learning frameworks like Keras and TensorFlow with GPU acceleration using free and open source software like Greenplum Database and the Apache MADlib machine learning library. Topics will include architecture, common usage patterns, scalability results and bright opportunities for the future.

Deep neural networks are very efficient at solving problems in domains such as computer vision, speech recognition and language translation. Once solely the purview of academia and Silicon Valley types of companies, deep learning is now making inroads in the enterprise by virtue of new algorithms, better tools, and lower costs for computation, storage and networking.

But enterprise data typically lives in relational and document form in databases, so how can you use this data for building deep learning models? You could try to move it out to a separate execution engine, but it is suboptimal to copy huge amounts of data between systems. What about the idea of building deep learning models directly in the database, bringing the compute to where the data lives?

It’s possible and I look forward to discussing this topic at FOSDEM’19!

As graph processing moves to the mainstream, a large number of specialized graph engines have emerged. However, for many enterprises, much of their important data resides in relational databases and SQL is the most common workload. So is it reasonable to suggest that relational data processing engines can be used to solve graph problems in a productive and performant manner?

The answer to this question is: “Yes!”

In this talk, we will address the use of massively parallel processing (MPP) databases for graph analytics workloads. We will share some recent findings from the Apache MADlib (incubating) project, including design of graph data structures, implementation of common graph algorithms, and performance results.

Graph analytics is becoming an important part of enterprise computing. With roots in academia going back many decades, the last 10-15 years have seen a huge surge of interest in this topic to address a wide range of modern use cases, from cybersecurity to social networks to supply distribution chains.

Enterprises have made significant investments in infrastructure, software, and training of their employees, all centered around SQL. So how can an enterprise add graph analytics to their business without the cost and complexity of moving to specialized graph processing engines? And, what are the tradeoffs?

Graph analytics is a new area of innovation in Apache MADlib, which is a SQL-based open source library for scalable in-database analytics. It provides parallel implementations of mathematical, statistical and machine learning methods for structured and unstructured data.

Many existing analytics products do not scale in a way that makes it convenient and economical to operate on large data sets. The graph methods in Apache MADlib have been designed to take advantage of the shared-nothing, scale-out parallelism offered by modern parallel database engines.

I look forward to presenting this topic at FOSDEM 17!

Apache MADlib (incubating) is an innovative SQL-based open source library for scalable in-database analytics. It provides parallel implementations of mathematical, statistical and machine learning methods for structured and unstructured data. MADlib also has an R interface for data scientists who prefer to work in R.

In this talk, we will describe the impetus behind creating a SQL-based scale-out machine learning project, review the architecture and implementation, and describe some of the recent functionality added by the Apache community. We will also present the R interface to MADlib, called PivotalR.

Finally, we will discuss the future direction of the project and invite big data developers and data scientists to participate in Apache MADlib, for both fun and profit.

The primary goal of the MADlib project is to accelerate innovation in the data science community via a shared library of scalable in-database analytics.

Many existing analytics products do not scale in a way that makes it convenient and economical to operate on very large data sets, which is a more and more common scenario in this era of Big Data. The methods in MADlib have been designed to take advantage of the shared-nothing, scale-out parallelism offered by modern parallel database engines.

Currently MADlib runs on the following open source platforms: Apache HAWQ (incubating) Hadoop-native SQL Database, Greenplum database, and PostgreSQL. Big Data has also brought about a renewed interest in query optimization, so as these platforms innovate in the area of distributed query performance, libraries such as MADlib can benefit greatly and provide even more significant benefits to research and commercial organizations that want to reason over very large data sets.

Some other key principles driving the architecture of MADlib are:

• Operate on the data locally in-database. Do not move data between multiple runtime environments unnecessarily.

• Utilize best of breed parallel database engines, but separate the machine learning logic from database specific implementation details.

• Foster open community development, including active ties to academic research.

We look forward to presenting this topic at FOSDEM’16!