Big data infrastructure management firm Pepperdata Inc. is turning its attention to high-performance workloads, launching a new monitoring tool today for graphics processing units that power data-intensive applications such as Apache Spark on Kubernetes.

Pepperdata sells software that’s used to manage and fine-tune workloads that run on big data frameworks such as Apache Hadoop and Spark. It does this by automatically scaling systems resources based on a deep understanding of how each application runs, using hundreds of real-time performance metrics. The company says this helps enterprises to squeeze the best possible performance from each app and track their resource spend for clear accountability.

Today’s new offering is designed to get the best possible performance out of a particular breed of workloads that harness “tremendous amounts of data,” the company said. By that, it means artificial intelligence and machine learning workloads that these days generally rely on GPUs for the higher performance they provide.

Pepperdata says the extreme processing power afforded by GPUs comes with a high price tag, which means enterprises can benefit from constant monitoring to ensure minimal resource waste and get the best performance at the lowest possible cost.

It said the new offering will enable customers to tweak and improve the performance of their Spark apps running on GPUs at a much more granular level.

Pepperdata’s secret is that it provides visibility into GPU resource usage at the application level, unlike traditional infrastructure monitoring tools that are limited to the platform level only. As well as monitoring their performance, Pepperdata said, it also provides recommendations on how to optimize GPUs to squeeze more juice out of them.

Pepperdata Chief Executive Ash Munshi said Apache Spark on Kubernetes has emerged as a key part of the infrastructure of thousands of AI and machine learning workloads.

“GPUs can handle these workloads but they are expensive to buy and they are power-intensive,” he said. “Up until now enterprises haven’t had a way to view and manage GPU infrastructure and applications, which leads to unnecessary waste and overspending on big data workloads.”

Image: Pepperdata