 Welcome back to theCUBE's coverage of ISC High Performance 2023, where we're covering all things HPC, we're talking machine learning, AI, high performance analytics, we got sessions on sustainability, a little bit of quantum, and one of our favorite topics, how the storage landscape in HPC is changing. With me are Anthony Dina, who's the Global Field CTO for unstructured data solutions at Dell Technologies and Shervin Samak, who is the HPC Product Manager at Dell. Guys, welcome. Good to have you on theCUBE. Thank you, Brian. Thank you so much, yeah. Shervin, maybe you could start us off, I mean, storage has gone from spinning mechanical rust in the 2010s to, I always say finally, you got machine speeds with flash, but how specifically is high performance storage in the 2020s different from what it looked like in the 2010s? Sure, yeah. So we are seeing different requirements and this requirement kind of enforced by the customer and their workload. Back in 2010, it was mostly HPC, lots of sequential IO, and now we are seeing the convergence of AI into HPC and that brings more challenge and the challenge going to be more IOPS, highest throughput for the performance storage. That being said, it brings more flash to storage. Back in 2010, again, it was based on spindles. We are seeing a lot of more flash, specifically on NVMe blended into HPC storage. That's one of the changes we have seen in the past 10, 15 years. The other thing is the introduction of cloud. So data protection is important and we are not seeing only the deployment of data just on-prem, but rather in a colo as well as public cloud. So a lot of object storage also introduced in this storage as well. So that's another thing we didn't see back in 2010. So that's interesting. I want to ask you guys later on, I want to ask you about the sort of changing storage hierarchy, but a lot more diversity is what I'm hearing. So Anthony, I wonder if you could comment. Think about large scale HPC been dominated by parallel file systems. I remember listening to way back in the early super computer days, how you can share data across multiple network nodes and in parallel. So it's perfect for data intensive workloads like now AI, especially for tightly coupled systems. My question is, do you see parallel file systems sort of maintaining their prominence and dominance in HPC in the future? Well, it's funny. I mean, there's so much work that's happening with scale out NAS for embarrassingly parallel applications. It's the MPI applications where we typically see the holding of physics or space where that many to one right to a single file or a file that extends beyond a cash tier where parallel files has become interest. But what's interesting as we look at the workflow, for example, in the research community or around healthcare life sciences, the workflow as the data originates through an NFS right to a shared storage system will eventually get analyzed, moving it into the HPC cluster, results sets return back to the same folder so that the scientists can focus on the science. And so it's really part of a continuous loop. And I think that's what's changing. So even though parallel file systems gets all of the shiny lights at a lot of these conferences, the real workhorse in many of the enterprise applications is not. So, okay, so is the implication on that? If I could just follow up, is that a benefit to workflow? Is it cost? Does it just mean you could do more work with less? It's about continuous operation. So instead of creating data, moving it into the cluster and then bringing it back, job schedulers that can move those relevant data sets back into a place where it gets the intensive computational analysis and then returned is what we're talking about. So it's a workflow optimization more than anything else. But that doesn't mean that parallel file systems are fading. In fact, as particularly as we look for these AI intensive throughput oriented workflows, it's still very, very relevant. Interesting, thank you. Shervin, I love talking about HPC because there's so many technology advancements in the space. I mean, obviously, the processor technology has been so much focused on and is diversifying, but you got improvements. We were just talking about parallel models. You got faster interconnects, optical, AI, you got to throw quantum in there into the mix, et cetera. From your point of view, what are the technologies that are sort of changing the game in these high performance file systems that have the greatest impact on performance and outcomes? Sure, so we can talk about it in maybe two different areas. One is on the hardware and the other one is basically on the software side of it, the standard protocol that had been introduced. On the hardware side, we saw an introduction of PCIe and that's kind of revolutionized storage specifically. The other one was NVME. So NVME has started to come in around like 2011 and in the past almost 14 years, we see the throughput with the NVME has been gone up twice, actually quadruple, 4X. So that's the hardware side of it, but when we come to the software side, we see different introduction of CXL, very important. We are going to see a lot more about it in the next coming years. RDMA will come with the internet. So it's helping storage as well. And I forgot one point about hardware and that's 3D NAND. So like NVME came in and then we see the QLC and TLC coming in to kind of bring down the price of the NVME. These are kind of the advantages that we have seen in the past few years, but we didn't have it back in 2010. Just to follow up on that, Sherman, if I could, it seems to me that applications in the 2010s were written assuming you had a slow spinning I.O. So I could go off and do other things waiting for that. And now you're talking about PCIe, NVME, 3D NAND, Flash, where you could do atomic writes, you're not having to wait for chatty protocols. How has that affected the software layer and the way in which people think about writing apps? So that's where like the file system is going to become very important. Like it's actually a different level of file. So we have the file system and on top of that, we have the application. Think about it, maybe TensorFlow or storage at that level, you have to interact with the file system. So we are seeing a lot of those applications are going to be tuned and optimized to interact with the file system. That's one of the areas. But we're also seeing a lot more maybe features and maybe updates coming into the file system to take advantage of the high throughput media like NVME. So two areas, application side, as well as the file system. Yeah, got it. And then I want to come back and follow up on the parallel file systems conversation we're having and maybe talk about are there possible alternatives? I mean, you hear about things like, you read about neuromorphic computing, quantum's in there. I've read about DNA computing and I've even read about this thing called approximate computing. I mean, I love back in napkin math so it sounds appealing to me unless it's like designing a safety tolerance of an elevator or a thousand other things. But at any rate, are there alternatives to parallel file systems emerging? And if so, where are they gaining traction? Is it in simulation, is AI changing things? Where is the potential if it even exists? Well, look, ultimately we're talking about maximizing someone's budget to getting the outcomes that they need through throughput capacity and available storage. And whether it's by design, whether it's because the hyperscale cloud providers have offered a new kind of throughput capabilities in the cloud, we see object more and more as a part of the overarching ecosystem. So it's less of a replacement and more of an augmentation but here is where we see the mullet opportunity. So instead of creating a large, flat, fast and cheap scratch space, wanna not have your parallel file system do what it does best and that is render the rights and have fast reads at a cash or NVMe tier and in the back you can create a large S3-compatible repository. And so we look at collapsing the three or four tiers between scratch, the home directories and the archive to maybe just two. So that you have all the performance writing on NVMe and then you treat the backend as a persistent and archival tier in one element. Are you saying you're using the object as a sort of a cheap and deep get put simplification tier or am I understanding that right? And you've got higher level sort of cash. Yeah, I mean, look, object has made a name for itself and it's ability to find small number of files very, very quickly. It's associated with inexpensive but what we've seen with all flash object and other increased use cases on cloud native applications, it's actually become a primary data store for many different applications. And so you can find that object could be of interest outside of an HPC and having a single repository in which you can interrogate makes life a ton easier. Yeah, interesting, definitely simplification move there. Okay, sure. So let's talk a little bit about the hierarchy. So there's a lot of talk about in-memory computing. We're always going to have a storage hierarchy. We were just talking about it collapsing a little bit. What's your view of what the storage hierarchy is going to look like within large scale HPC? Will it collapse? Will it become more granular with non-volatile buffers in between memory and other storage layers? Whether it's primary, we talked about deep archive, remote vaults, the cloud, maybe I just answered my question. Is it getting simpler or more granular? Paint a picture for us. I think it's going to stay the same actually, the same level of tiers that we kind of defined in the storage. It's going to be there maybe like the share of each one goes toward the other. But in reality, just because of the dollar to terabyte or different feature of each storage tier, I think we are going to see different levels or tiers of storage stay. So parallel file system, as we talked about it, it's going to be there. Scale out NAS that's provides the data protection, all the things, some of those feature, it's going to be there as well as whatever goes to the cloud. So I think it really depends on what customer really wants and depends on some of the requirement, either it's the price, either it's some of the functionality, data management, everyone is going to pick and choose and create a solution around what kind of address their problem. I would add on top of that, honestly, the linchpin in the conversation isn't necessarily the throughput as much as it's the data management, knowing exactly what you have, feeding these exabyte scale requirements implies a level of control of being able to move those assets around. And so data management is going to be a very interesting function that may have been a little bit of lower priority. Follow up on that, when you say data management, you mean sort of knowing where stuff is and being able to put it on the right tier or is it more data coherence? What's the outcome of better data management? Well, certainly the cost to serve and be able to move it in the right place and location, but it's even as simple as, what was the result sets from the study I ran? When did I run it? Who owns it? Just the findability in asset tagging or metadata tagging is really important. And that will become increasingly important as the volume increases. Got it. So I was having a conversation today with somebody we were talking about, okay, what's the right mix for HPC? Should it be on-prem? Should the cloud guys are doing their thing? And as always, and it depends. If you're doing experimentation, something quick in the cloud makes sense to rent. The other factor was, hey, if you want to show off university and your high performance computing center, obviously it makes sense to have your own data center. There's like, hey, if you're going to be 50,000 cores, you want to have that to be sort of a non-prem cluster. But what does the hybrid multi-cloud HPC storage architecture look like? And not just the HPC storage architecture within the system, but within a customer's data center, across that customer's data center, out to any systems where their data resides, colos, clouds. What do you guys think about that? I think that data movement piece that Anthony was talking about is to become, again, the most important item here, even for the cloud hybrid environment, whatever it is. So because at the end of the day, customers wanted to have data somewhere, if they are putting it somewhere else, there needs to be some sass layer to take care of the data management for them. So, or if they have like other researchers or other maybe entity in the same organization wanted to access that data, how are we doing this data movement? So then it becomes the issue of the collaboration and then again, come back to the data management. That's my take on this overall hybrid plot on prem plus colo. Thoughts on that, Anthony? For me, it's all about the, and I'm gonna use the word loosely business demands. So look, if you've spent your money but an urgent project needs to get run, you are likely to go to the quote unquote spot market and use cloud resource to get that job done, even though it may be a more expensive value proposition, time to results matter. So that's one attribute. The second is, look, even some of the best institutions with the greatest scientists resident can run out of capacity. And so we're seeing increased levels of collaboration across the university space, which then implies some kind of geo-distributed way of allowing that information. In fact, it's not just data that's being generated inside of those institutions that matters, it's publicly available information. So the multicloud hybrid cloud function is an important attribute to think about, what is the data flow between the owners and how can the underlying architecture support those outcomes? I'm envisioning a supercomputer super cloud. Guys, really compelling discussion. I mean, as always on this great topic, I really appreciate your insights today. Thank you very much. You're welcome. And thank you for watching our continuous coverage of ISC 23 and the innovations in high performance computing. You're watching theCUBE, your leader in enterprise and emerging tech coverage.