Case Study: Dyna Robotics achieves faster, multi-cloud AI training with Alluxio

Dyna Robotics eliminates 30%+ training slowdowns with Alluxio

Dyna Robotics, a company developing embodied AI systems, faced significant performance bottlenecks when training its foundation models. Each training job required accessing tens of thousands of small files from Google Cloud Storage (GCS), which created high latency and slowed down their NVIDIA H100 GPU clusters. Their initial workaround, a self-managed NFS layer, introduced operational complexity, bandwidth contention, and a 30% performance degradation under heavy concurrency. To eliminate these I/O bottlenecks without adding more storage infrastructure, Dyna Robotics implemented a software solution from Alluxio.

Alluxio provided a unified data access layer by deploying its distributed caching software on Dyna's GPU nodes. This solution pooled the local SSDs into an 88TB global cache, presenting a POSIX interface to the training code for transparent integration. The results were immediate: Alluxio eliminated the periodic 30%+ training slowdowns, removed operational complexity from manual data management, and enabled a true multi-cloud architecture. This allowed Dyna to seamlessly run workloads across GCP and Together AI without re-ingesting data or incurring cross-cloud egress fees, transforming their infrastructure into a scalable and portable AI data platform.

Dyna Robotics