AWS HPC Blog

Category: Artificial Intelligence

Enhancing ML workflows with AWS ParallelCluster and Amazon EC2 Capacity Blocks for ML

Enhancing ML workflows with AWS ParallelCluster and Amazon EC2 Capacity Blocks for ML

No more guessing if GPU capacity will be available when you launch ML jobs! EC2 Capacity Blocks for ML let you lock in GPU reservations so you can start tasks on time. Learn how to integrate Caacity Blocks into AWS ParallelCluster to optimize your workflow in our latest technical blog post.

Improving NFL player health using machine learning with AWS Batch

Improving NFL player health using machine learning with AWS Batch

In this post we’ll show you how the NFL used AWS to scale their ML workloads and produce the first comprehensive dataset of helmet impacts across multiple NFL seasons. They were able to reduce manual labor by 90% and the results beats human labelers in accuracy by 12%!

Figure 2: Identification of redun jobs and grouping them into Array Jobs to run on AWS Batch. (Top) redun represents the workflow as an Expression Graph (top-left), and identifies reductions (red boxes) that are ready to be executed. The redun Scheduler creates a redun Job (J1, J2, J3) for each reduction and dispatches those jobs to Executors based on the task-specific configuration. The Batch Executor allows jobs to accumulate for up to three seconds (default) in order to identify compatible jobs for grouping into an Array Job, which are then submitted to AWS Batch (top-right). (Bottom) As jobs complete in AWS Batch, the success (green) and failure (red) is propagated back to Executors, the Scheduler, and eventually substituted back into the Expression Graph (bottom-left).

Data Science workflows at insitro: how redun uses the advanced service features from AWS Batch and AWS Glue

Matt Rasmussen, VP of Software Engineering at insitro, expands on his first post on redun, insitro’s data science tool for bioinformatics, to describe how redun makes use of advanced AWS features. Specifically, Matt describes how AWS Batch’s Array Jobs is used to support workflows with large fan-out, and how AWS Glue’s DynamicFrame is used to run computationally heterogenous workflows with different back-end needs such as Spark, all in the same workflow definition.

Figure 1: Evaluating a sequence alignment workflow using graph reduction.** In redun, workflows are represented as an Expression Graph (left) which contain concrete value nodes (grey) and Expression nodes (blue). The redun scheduler identifies tasks that are ready to execute by finding subtrees that can be reduced (red boxes), substituting task results back into the Expression Graph (red arrows). The scheduler continues to find reductions until the Expression Graph reduces to a single concrete value (grey box, far right). If any reduction has been done before (determine by comparing an Expression's hash), the redun scheduler can replay the reduction from a central cache and skip task re-execution.

Data Science workflows at insitro: using redun on AWS Batch

Matt Rasmussen, VP of Software Engineering at insitro describes their recently released, open-source data science framework, redun, which allows data scientists to define complex scientific workflows that scale from their laptop to large-scale distributed runs on serverless platforms like AWS Batch and AWS Glue. I this post, Matt shows how redun lends itself to Bioinformatics workflows which typically involve wrapping Unix-based programs that require file staging to and from object storage. In the next blog post, Matt describes how redun scales to large and heterogenous workflows by leveraging AWS Batch features such as Array Jobs and AWS Glue features such as Glue DynamicFrame.

Scalable and Cost-Effective Batch Processing for ML workloads with AWS Batch and Amazon FSx

Batch processing is a common need across varied machine learning use cases such as video production, financial modeling, drug discovery, or genomic research. The elasticity of the cloud provides efficient ways to scale and simplify batch processing workloads while cutting costs. In this post, you’ll learn a scalable and cost-effective approach to configure AWS Batch Array jobs to process datasets that are stored on Amazon S3 and presented to compute instances with Amazon FSx for Lustre.