Auto-Tiered Storage for AI and HPC

DataDirect Networks (DDN) has launched EXA5, the company’s fifth-generation Exascaler Lustre file system platform, which will be used to populate the company’s all-flash, mid-range, and high-end storage appliances. As is DDN’s custom, Exascaler is aimed at the HPC crowd, but this latest release also contains a number of features designed to attract more enterprise-oriented customers, especially those doing AI work.

“It combines the capabilities of our storage, which have always historically been focused on exceptional performance, with some enterprise features as well, as these demands become more and more apparent for those types of organizations that aren’t as familiar with DDN,” said Kurt Kuckein, DDN’s senior marketing director.

The company has been methodically adding enterprise-friendly features to the Exascaler software for years, but the AI market is spurring DDN to give this area extra attention. For EXA5, a lot of these new features are fairly standard issue – enhanced data protection, upgraded security (including for multi-tenancy), support for non-Lustre files via CIFS and NFS gateways, and auditing capabilities. Improved small file performance was also added, although nowadays that’s also appreciated by traditional HPC customers as well.

Kuckein told us they did a lot of work accelerating small file performance in Lustre. At the hardware level, this is primary made possible by tapping into the exceptional random access performance of flash-based storage components. Kuckein says the optimization they’ve done not only speeds I/O for small files, but also across a range of files sizes, from a few kilobytes to multiple megabytes.

However, the big new EXA5 feature is STRATAGEM, a brand new policy engine that provides automated tiering between flash and hard disk. The idea behind tiering is to optimize I/O performance by keeping active data in flash and placing data accessed less frequently on spinning disks. It does this by scanning the file system in the background and moving files around if the access patterns warrant it.

Up until now DDN customers have relied on open source Lustre policy engines, like Robinhood, which according to Kuckein, left something to be desired, performance-wise. Strategem, he said, uses much less overhead so that file system scanning can occur without unduly impacting user performance. One interesting feature is that it keeps most flash-based files mirrored on hard disk in case it quickly needs space on the flash tier. Users can also develop custom policies to automate data promotion and demotions for specific situations. An API is also available to exert control over data movement driven externally, such as from the scheduler.

DDN has set up the policy engine so it doesn’t matter if the flash sits in its own NVME-based storage array, like the ES200NV or ES400NV, or a mixed with hard disks in a hybrid appliance, like the ES7990. It will figure all that out whether or not the flash is in the file namespace or the system itself.

EXA5 has already found a home on Frontera, the new supercomputer recently launched at the Texas Advanced Computing Centre (TACC). As the number five system in the world, powered by both CPUs and GPUs, the new TACC machine is expected to run AI workloads alongside conventional HPC applications. It’s backed 50 petabytes of DDN hardware, three petabytes of which in the form of NVMe-based hardware for extra speedy scratch storage.

In the AI space, the kind of customer DDN has in mind for EXA5 is XXII, a company that offers real-time video analysis software. It can be used for frictionless (cashless) retail environment to monitor store customers and tally up charges based on what they put in their cart. The XXII analytics software was demonstrated at the recent GTC conference, running on Nvidia hardware hooked up to DDN storage.

EXA5 can run atop the all-flash SFA200NV and SFA400NV, the hybrid SFA7990, and the high-density SFA18K, as well as the A³I storage appliances for AI (A1200 and AI7990). It will be available early in the third quarter of the year.

Sign up to our Newsletter

Featuring highlights, analysis, and stories from the week directly from us to your inbox with nothing in between.
Subscribe now

Be the first to comment

Leave a Reply

Your email address will not be published.


This site uses Akismet to reduce spam. Learn how your comment data is processed.