SC20 Showcases DDN’s Evolution towards HPC-plus-Enterprise Storage Power

https://www.hpcwire.com/2020/11/16/sc20-showcases-ddns-evolution-towards-hpc-plus-enterprise-storage-power/?fbclid=IwAR0E-LMBKVq4vFMyqLsVm3KNsrTYN6CYNPZnWfl0Hv01UE2EhbJbpeVtkME

Currently, there’s a lot on DDN’s plate as the long-time leader in HPC storage integrates recent acquisitions and strives to become a comprehensive HPC-plus-enterprise storage technology supplier. SC20 is providing a showcase for those efforts as DDN rolls out product updates, impressive supercomputer wins including Fugaku (fastest in the world), an expanded user forum and portal, and affirms its intention to create an enterprise-friendly version of Lustre under the EXAScaler product line.

(DDN had already begun winding down its GPFS-based (IBM Spectrum Scale) GRIDScaler offerings.)

The product news – mostly enhancements of EXAScaler 5 (v 5.2), DataFlow, and IntelliFlash – all emphasize enterprise, cloud, and AI features. Taken together with the supercomputer wins, DDN seeks to reinforce its ongoing commitment to the high-end HPC market and while simultaneously ramping up its ambitious expansion into the enterprise market.

The Fugaku win certainly helps the first cause. “We haven’t been the storage [choice] for the number one Top500 supercomputer in a while so it’s nice to get back into that position by supporting a system like Fugaku,” said Kurt Kuckein, DDN VP of marketing, in a pre-briefing with HPCwire. “We’re winning these types of large deals elsewhere,” he noted, pointing out that DDN is still the only certified vendor for Nvidia DGX SuperPOD.

Fugaku will use DDN SFA18K (block storage, combining hard disk drives and NVMe flash) for its second-tier storage system which requires a “total effective capacity” of at least 150 PB with effective throughput of 1.5 TB/s according to DDN. “Joint testing between DDN Japan and Fujitsu confirmed that a configuration of five SFA18KE and twenty-five SFA18KXE devices, mounted with Fujitsu’s FEFS file system, [successfully] met these requirements,” reports DDN.

Fugaku’s second-tier storage will be used as compute nodes for OSS (object storage servers) and serve as a gateway to data storage units. DDN says the “E” in SFA18KE and SFA18KXE stands for “embedded,” with OSS implemented on a virtual machine (VM) on the storage controller. Hard drives in expansion disk enclosures will serve as OSTs (object storage targets) used for data storage.

Another significant DDN win announced during SC20 is the University of Florida’s flagship AI supercomputer ‘HiPerGator’, which will use 10 DDN AI400X all-flash appliances in conjunction with 140 Nvidia DGX A100 systems at the core of HiPerGator.

Pivoting to the enterprise, Kuckein said “We have essentially three product announcements at SC20. 1) We’re releasing the next version of EXAScaler (5.2) building on the capabilities that we introduced in EXA5 (DDN parlance) when it was launched at ISC 2019. 2) DataFlow (data migration tool) has some increased capabilities, especially in terms of managing large volumes of data with EXAScaler and integration into the cloud. 3) We’re also introducing Tintri’s IntelliFlash product line into the DDN portfolio for some mid-range capabilities that would complement the existing DDN EXAScaler and A3I products.”

One of DDN’s challenges is bringing its expanding portfolio to market; it has chosen to market the high-end HPC products under the DDN At Scale brand while using the Tintri brand for its more mainstream enterprise offerings. IntelliFlash – a hybrid product that combines NVMe flash memory with HDDs for scalable capacity up to 5 petabytes – will be supported by both organizations.

EXAScaler 5.2 is the most relevant HPC product among the announcements. It is DDN’s Lustre-based (parallel file system) platform and is now being position not only for DDN’s traditional HPC customer base but also for enterprise HPC and more mainstream enterprise applications. Lustre, of course, has long had a reputation for being powerful but also tricky to work with and lacking in popular enterprise features. Since acquiring Intel’s Lustre division, DDN has been working to change the reality and perception.

“I don’t love the word, but the ‘democratization’ of the high-performance parallel file system to take it into more standard IT markets and making it friendly to those markets is exactly what we’re doing,” said Kuckein. “So EXA5 has been around just about 18 months and the core part of it was the stratagem policy engine and scanning engines that we introduced at that time.”

DDN has been layering features on top of the policy engine including for example, hot pools for easily moving data between tier as needed based on policy or individual job analysis. Those features got a boost in the most recent EXA5 release.

“We’re now also introducing search, sync, and snapshot, providing additional enterprise data protection management capabilities,” Kuckein said. “The file system search capability allows administrators refine their hot pools policies to visualize what’s going on within the file systems, identify places where they could be more efficient, identify trends, and figure out for the future what type of growth should be expected. The new synch capability has built-in backup and data sharing capabilities. It adds the ability to protect the file system with an off-site copy remotely, or it can be used also to move data copies to the cloud.”

The DataFlow product “is our external data management tool for data migration or data protection. It has robust backup archive capabilities to move data out of the primary file system and into a new file system structure for true archive. You can do incremental backups to a tape system or off the cloud or to another file system using this software.”

A major use case, perhaps not surprisingly, is migrating data from DDN’s installed base of GRIDScaler customers. GRIDScaler was based on IBM’s Spectrum Scaler (GPFS) file system.

“We bought the Lustre development team from Intel and really see that as our ability to control our own destiny. With Spectrum Scale, we had input to the road map but not control [over] what IBM chose to develop. [Dropping GRIDScaler] was a strategic decision to promote our own products ahead of someone else’s. That was one reason. Number two was IBM’s decision to change its licensing. Moving from a per core licensing to a capacity licensing didn’t make sense to us and the customer base [which] was ‘at scale’ customers who anticipate growing their storage capacities in vast quantities.”

It will be interesting to watch how the enterprise Lustre gambit plays out. Even within traditional HPC, many of GPFS’s popular ‘enterprise features’ such as snapshots and deduping have been desirable. DDN’s efforts along those lines may be welcome by many.

Incorporating the IntelliFlash product from Tintri into the DDN line adds a mid-range offering that is well-suited for handling structured data and more familiar enterprise file systems such as NFS.

“We see an enterprise AI use case for customers who are interested in DDN because of our expertise in dealing with at scale problems but have a more structured data problem. DDN’s A3I solution is [best] for unstructured data at really big scales – use cases like autonomous driving, genomics, natural language processing, image inference that require scalable file systems,” said Kuckein.

“But you have this other set of use cases for AI that we would historically consider smaller volumes of data, not necessarily part of the HPC world, but which need addressing and also need the expertise of folks who’ve been dealing with problems on the AI side. Their use cases tend to be around structured data sets, things like consumer preference analysis and credit card processing,” said Kuckein.

On the flip side, Kuckein said IntelliFlash has a fit within DDN customer base for uses such traditional user file sharing. “Home directories work quite well with IntelliFlash and they’ve got capabilities that we don’t have within EXA5 like de-duplication that customers are looking for in mid-range HPC and enterprise functions. For customers who are really dedicated to NFS and don’t want to explore the parallel file system, IntelliFlash is probably sufficient.’

He also noted that for customers whose system “have maybe eight or 16, GPUs, for instance, with one or two DGX systems, NFS is going to be okay, and if they don’t anticipate scaling beyond those needs, something like IntelliFlash would meet their requirements quite well.”