These are exciting times for those of us working on high performance computing (HPC) initiatives, and LUG 2019 had it all on display. In this brief blog post, we describe the highlights of the event and what they mean to organizations looking to do more with HPC.
Lustre Has Serious Momentum as HPC Usage Expands
While Lustre is already the dominant filesystem at big entities, many are expanding their use of Lustre on multiple fronts: increasing business opportunities, researching the next big thing, and providing access to data for multiple business and research use cases.
One example of Lustre’s expanding applicability came from ExxonMobil and their Lustre presentation at the conference. ExxonMobil’s Lustre implementation manages over 50 petabytes of data across an offsite and campus data center to fuel a wide range of HPC applications in the realms of molecular and quantum modeling, geophysics, drilling and subsurface modeling, and even facilities tracking. The presentation also touched on a trend in which large organizations are re-adopting or re-committing to Lustre. For their part, ExxonMobil initially adopted Lustre in 2008, switched to GPFS in 2009, then switched back to Lustre in 2012 (now on their fifth version – v18.104.22.168 with PFL).
Further growth of Lustre can be seen in the numerous efforts to make Lustre a default part of the Linux kernel. This LUG conference presentation, given by Neil Brown, Kernel Engineer at SUSE, makes the key point that incorporating Lustre into Linux will make Linux stronger. Of course, it will also serve to further expand Lustre to an even more dominant share of the HPC market.
Lustre Gains Applicability Beyond HPC
There was a lot of buzz at LUG 2019 about the new ways in which organizations are applying Lustre to modern applications that are not necessarily HPC. This presentation by NERSC shows how AI-driven apps can benefit greatly from Lustre, and this spurred a lot of discussion around the use of Lustre for AI, ML, EDA, and simulations taking place in the cloud (more on cloud below).
Optimizing Storage Costs is the Primary Concern for Lustre Users
More than 75 percent of all the LUG sessions focused on ways to optimize costs surrounding the use of Lustre, and storage costs were top of mind for most. The use of all-flash storage is certainly proving to be one effective way of containing storage costs, as modeled in detail in this NERSC presentation. Needless to say, much of the cost-containment discussion moved down the path toward the public cloud. Lustre on public cloud is now viewed as a great option to reduce cost of doing HPC, especially in areas such as EDA and ML.
Lustre in the Cloud is an Emerging Reality
By now, you have likely heard that AWS has been offering AWS FSx Lustre as a result of customers specifically asking for Lustre in the cloud. According to an AWS presentation, customers that need hundreds of GB/second of throughput and sub-millisecond latencies have been asking AWS for Lustre by name, recognizing that it is the only file system that can pull of consistently high performance on cloud. Having built our own multi-cloud data orchestration platform on top of Lustre, we recognize the superiority already and offer optimal cloud data movement performance to all our customers and prospects. It’s a large part of the reason Kmesh was recently named a TiE50 Award recipient.
Tuning Lustre to Cloud Requires Specialized Expertise
If making Lustre work on cloud in a highly performant manner were easy to do, you would already see multiple cloud platforms with FSx Lustre offerings that work. But you don’t see that today. Our own Kmesh presentation at LUG explains why that is. The fact is, Lustre on cloud is tough to do, and there are lots of variables to consider. Not to toot our own horn, but we have done it, which is why Kmesh is leading the charge to help others do it as well.
To help the HPC community and to expand the use of Lustre in the cloud, Kmesh has already performed some work on HPC benchmark io500 on cloud, and we continue to engage with io500 leadership in an effort to ensure io500 incorporates cloud elements into their benchmark.
As a result of our ongoing interactions with io500, Kmesh co-founder Vinay Gaonkar has been invited to present at both io500 and HPC BOF sessions at ISC in June. According to ISC, The ISC High Performance conference will bring together over 3,500 researchers and commercial users of high performance computing, networking, and storage. You can visit them on the web at https://www.isc-hpc.com.
Video Series: Why Lustre?
Check out videos about why leading HPC enterprises choose Lustre for their computing needs. VIDEOS