Performance and Ease of Use Enhancements for GPU-Centric AI/ML Workloads

New features streamline data pre-processing and loading phases to enable better use of GPUs to improve AI/ML training efficiency while reducing cost.


I've had the opportunity to meet with Haoyuan Li, Founder and CEO of Alluxio several times. Alluxio, the open-source data orchestration software for large-scale workloads, has just released version 2.6 of their Data Orchestration Platform. The new release features enhanced system architecture enabling AI/ML platform teams using GPUs to accelerate their data pipelines for business intelligence, applied machine learning, and model training.

“Enterprises seeking a competitive advantage are making greater use of machine learning and AI to derive insights from massive datasets,” said Li. “These datasets are often distributed across hybrid cloud environments, making more consistent and efficient data access critical to realizing the value from their AI/ML initiatives.”

In the latest release, Alluxio improves its system architecture to support AI/ML applications using the POSIX interface. System performance is maximized by removing inter-process latency overheads, which is critical for enabling full utilization of compute resources. Aside from I/O performance, the end-to-end workflow of data preprocessing, loading, training, and result writing are supported by Alluxio’s data management capabilities.

“Machine learning applications benefit greatly from the performance acceleration offered by GPUs. However, when utilizing powerful compute hardware, the limiting factor of the workload often shifts to I/O where workloads become bound on how fast data can be made available to the GPUs as opposed to how fast the GPUs can do training computations,” said Adit Madan Product Manager, Alluxio. “Alluxio 2.6 bridges this gap in performance with a data orchestration layer for AI/ML workloads, allowing applications to fully utilize expensive and powerful hardware without encountering the data access and I/O bottlenecks.”

Alluxio 2.6 Community and Enterprise Edition features new capabilities, including:

Faster Data Access with a Large Number of Small Files

Alluxio 2.6 unifies the Alluxio worker and FUSE process. By coupling the two, significant performance improvements are achieved due to reductions in inter-process communication. This is especially evident in AI/ML workloads where file sizes are small and RPC overheads make up a significant portion of the I/O time. In addition, containing both components in a single process greatly improves the deployment of the software in containerized environments like Kubernetes. These enhancements substantially reduce data access latency, enabling users to process greater amounts of data more efficiently to deliver more AI/ML benefits to the business.

Simplified Data Management and Operability

Alluxio 2.6 enhances the mechanism to load data into Alluxio managed storage and introduces more traceability and metrics for easier operability. This distributed load operation is a key portion of the AI/ML workflow, and adjustments to the internal mechanisms have been made to optimize for the common case of loading prepared data for model training.

Improved System Visibility and Control

Alluxio 2.6 adds a large set of metrics and traceability features enabling users to drill into the system’s operating state. These range from aggregated throughput of the system to summarized metadata latency when serving client requests. This new level of visibility can be used to measure the current serving capacity of the system and identify potential resource bottlenecks. Request level tracing and timing information can also be obtained for deep performance analysis. These new features enable users to get new levels of visibility and control for improving SLAs of their large-scale data pipelines for a wide variety of use cases.