

Alluxio Inc., which sells a commercial version of an open-source distributed filesystem and cache, today announced new features that accelerate artificial intelligence model training and enhance integration with Python software development kits.
The company said the updates collectively enable organizations to train models faster, handle large datasets more efficiently and simplify complex AI infrastructure.
Alluxio said enhancements are intended to support fast, prioritized access to important training data and integrate with common AI frameworks. The company has pivoted to address AI model training, a process that can take months, with promises of significant performance improvements. “We see DeepSeek as an opportunity,” said founder and Chief Executive Haoyuan Li, referring to the Chinese startup that tanked tech stocks this week with news of its low-cost approach to model training. “It creates an easier sell for us.”
Last July, the company trumpeted enhancements that it said can improve utilization of costly graphics processing units to 97%. “Everybody’s running very fast to take advantage of AI, so we help them innovate faster, accelerating training workloads, getting models into market faster, learning how they’re being used, and bringing that info back into the model training process,” said Bill Hodak, vice president of marketing and product marketing. “The faster they can do that, the more advanced and accurate their models will be.”
Alluxio Enterprise AI version 3.5 includes an experimental CACHE_ONLY write mode that the company said significantly improves the performance of write operations. When enabled, it mode writes data exclusively to the Alluxio cache instead of the underlying file system, eliminating bottlenecks associated with storage systems.
Hodak said the feature is particularly useful with checkpoint files, which are saved snapshots of a model’s state at a given point that can be used to resume from a saved point rather than restarting from scratch. Hodak said that the files can be large and cause long delays in the training process while loading. “If it was taking an hour before, it probably takes 20 minutes now,” he said.
Advanced cache eviction allows administrators to enforce time-to-live settings on cached data, which define how long cached data remains valid before it is automatically expired and removed. Administrators can now define caching priorities for specific data that override Alluxio’s default.
“least recently used” algorithm to keep data in the cache that would otherwise be expunged.
“The goal is to reduce as much overhead as possible,” Hodak said. “This improves cache hit ratios, which depends on the workload.”
Another experimental feature is enhanced integration between Alluxio’s Python SDK and popular AI frameworks like PyTorch, PyArrow and Ray. The integrations provide a unified Python filesystem interface, enabling applications to interact seamlessly with local and remote storage systems.
The release also introduces several enhancements to Alluxio’s application programming interface for accessing data in S3 object storage.
Support for HTTP persistent connections maintains a single TCP connection for multiple requests. This reduces the overhead of opening new connections for each request and decreases latency by approximately 40% for 4KB S3 ReadObject requests, the company said.
Communication between the Alluxio S3 API and the Alluxio worker now supports TLS encryption and multipart upload. The latter splits files into multiple parts for faster parallel uploads.
Hodak said that a new caching service improves the performance of very large directory listings, serving up results up to five times faster by delivering directory listing metadata from the cache and speeding performance up to five times.
Administrators can now set a rate limit to control the maximum bandwidth an individual Alluxio Worker can read from the Under File System, the underlying storage system Alluxio uses to store data for cache access.
Clusters can now have worker nodes with heterogeneous CPU, memory, disk and network configurations, enhancing flexibility.
THANK YOU