 From the SiliconANGLE Media Office in Boston, Massachusetts, it's theCUBE. Now, here's your host, Stu Miniman. Hi, I'm Stu Miniman, and welcome to a special CUBE conversation here in our Boston area studio. Happy to welcome back to the program, Danny Cobb, who's with Dell EMC in the CTO office. Hey Stu, great to see you here today. Great to see you too. So Danny, we're going to talk about a topic that, like many things in the industry, it seems like it's something that happened overnight, but there's been a lot of hard work going on for quite a lot of years. Even going back to, heck, when you and I worked together, company used to be called EMC, NVME. So first of all, just bring everybody up to speed as to what you work on inside the Dell family. Sure, so my responsibility at now Dell EMC has been this whole notion of emergent systems, new technologies, new capabilities that are just coming into broad market adoption, broad readiness, technological feasibility and those kinds of things. And then making sure that as a company, we're prepared for their adoption and inclusion in our product portfolio. So it's a great set of capabilities, a great, great set of work to be doing, especially if you have a short attention span like I do. Great, so Danny, I spent a lot of time these days in the open source world. You talk about people are moving faster, people are trying lots of technologies. You've been doing some really hard work, you know, the company and the industry in the standards world. So maybe just tell, what's the importance of standards these days and, you know, bring us back to how this NVME stuff started. Yeah, so great way to get everybody up to speed on this. You know, as you mentioned when you kicked off, you know, NVME and overnight success, almost 11 years in the making now. The very first NVME standard was about 2007. EMC joined the NVME consortium in 2008 along with an Austin, Texas computer company called Dell. So Dell and EMC were both in the front row of defining the NVME standard and essentially putting in place a set of standards, a set of architectures, a set of protocols, product adoption capabilities, compatibility capabilities for the entire industry to follow starting in 2008. Now you know from our work together that the storage industry likes to make sure that everything's mature, everything works reliably, everything has broad interoperability standards and things like that. So since 2008, we've largely been about how do we continue to build momentum and generate support for a new storage technology that's based on broadly accepted industry standards in order to allow the entire industry to move forward, not just to achieve the most out of the flash revolution but prepare the industry for coming enhancements in storage class memory. Yeah, so storage class memory, you mentioned things like flash. One thing we've looked at for a long time is when flash rolled out, there was a lot of adoption on the consumer side first and then that drove the enterprise piece but flash today is still done through a SCSI interface with SAS or SATA and believe we're finally getting rid of, when we go to NVMe, what some of the industry have called the horrible SCSI stack. That's right. So explain to us a little bit about first, the consumer piece of where this fits first and how it gets in the enterprise, where are we in the industry today with us? Yeah, so as you pointed out, a number of the new media technologies have actually gained abroad acceptance and a groundswell of support starting in the consumer space. The rapid adoption of mobile devices whether initially iPods and iPhones and things like that, tablets where the more memory you have, the more songs you carry, the more pictures you can take, a lot of very virtuous things, virtuous cycle type things occurred in the consumer space to allow flash to go from a fairly expensive, perhaps niche technology to broad, high volume manufacturing and with high volume manufacturing comes much lower cost. And so we always knew that flash was fast when we first started working out at EMC in 2005. It became fast and robust when we shipped in 2008. It went from flash to robust to affordable with technologies like the move from SLC to MLC and now TLC flash and the continuing advances of Moore's Law. And so flash has been the beneficiary of high volume consumer economics along with our friend Moore's Law over a number of years. Okay, so on the envy of me piece, your friends down in Round Rock and Dell, they've got not only the storage portfolio but on the consumer side, there's pieces like my understanding, NVMe is already in the market for some parts of this today, correct? That's right. I think one of the very first adoption scenarios for NVMe was in a lightweight laptop device. The storage stack could be more efficient. The fundamental number of gates in silicon required to implement the stack was more efficient. Power was more efficient. So a whole bunch of things that were beneficial to a mobile high volume client device like an ultralight or ultra portable laptop were made at a great place to launch the technology. Okay, and so bring us, what does that mean then for storage? Is that available in kind of the enterprise storage today and where is that today and where are we going to see in the next year or so? So here's the progression that the industry has more or less followed. We went from that high volume ultralight laptop device to very inexpensive M.2 devices that could be used in laptops and desktops more broadly, also gained a fair amount of traction with certain use cases and hyperscalers. And then as the spec matured and as the enterprise ecosystem around it, broader data integrity type solutions in the silicon itself, a number of other things that are bread and butter for enterprise class devices. As those began to emerge, we've now seen NVMe move forward from laptop and client devices to high volume M.2 devices to full function, full capability, dual ported enterprise NVMe devices really crossing over this year. Okay, so that means we're gonna see not only in the consumer pieces but should be seeing really enterprise rollout in, I'm assuming things like storage arrays, maybe hyperconverged, all the different flavors in the not too distant future. Absolutely right. The people who get paid to forecast these kinds of things when they look into their crystal balls, they've talked about when does NVMe kind of get close enough to its predecessor, SAS, to make the switchover be a no brainer. And oftentimes you get a performance factor where there's more value or you get a cost factor where suddenly that becomes the way the game is won. In the case of NVMe versus SAS, both of those situations, value and cost are more or less a wash right now across the industry. And so there are very few impediments to adoption. Much like a few years ago, there were very few impediments to adoption of enterprise SSDs versus high performance HDDs, the 15Ks and the 10K HDDs. Once we got to close enough in terms of cost parity, the entire industry went all flash overnight. Yeah, it's a little bit different than say the original adoption of flash versus HDD, HDD versus SSD. Remember back you had to have like the algebra sheet and you said, okay, how many devices did I have? What's the power savings that I could get out of that? Plus the performance that I had and then does this make sense? It seems like this is a much more broadly applicable type of solution that we'll see a much faster adoption. Do you remember those days of a little goes a long way? And then more is better and then all must be really good. And so that's kind of where we've come over what seems like a very few years. Okay, so we've only been talking about NVMe. The thing, I know David Floyer has been looking a lot from an architectural standpoint where we see benefit obviously from NVMe but NVMe over fabrics is the thing that has him really excited if you talk about the architecture. Maybe just explain a little bit about what I get with NVMe and what I'll get added on top with the over fabric piece of that and what's that kind of rollout look like? Can I tell you a little story about the, what I think of as the birth of NVMe over fabrics? Please. So some of your viewers might remember a project at EMC called Thunder. And Thunder was PCIE Flash with an RDMA over Ethernet front end on it. We took that system to Intel Developers Forum as a proof of concept. Around the corner for me was an engineer named Dave Mintern who's an Intel engineer who had almost exactly the same software stack up and running except it was an Intel RDMA capable NIC and an Intel flash drive and of course some changes to the Intel processor stack to support the use case that he had in mind. And we started talking and we realized that we were both counting the number of instructions from packet arriving across the network to bytes being read or written on this very fast PCIE device and we realized that there has to be a better way. And so from that day, I think it was September 2013, maybe it was August, we actually started working together on how can we take the benefits of the NVMe standard that exists mapped on to PCIE and then map those same primitives as cleanly as we possibly can on to at that time Ethernet but also Infiniband, Fiverr Channel and perhaps some other transports as a way to get the benefits of the NVMe software stack and build on top of the new high performance capabilities of these new RDMA capable interconnects. So it goes way back to 2013. We moved it into the NVMe standard as a proposal in 2014 and again, three, four years later now we're starting to see solutions roll out that begin to show the promise that we saw way back then. Yeah, and challenge with networking obviously is, sounds like you've got a few different transport layers that I can use there, probably a number of different providers, how baked is the standard? Where do things like, I hate to say interoperability fit into the mix? When do customers get their hands on it and what can they expect the roll out to be? So we're clearly at the beginning of what's about to be a very, I think, long, healthy future for NVMe over fabrics. I don't know about you, I was at Flash Memory Summit back in August in Santa Clara and there were a number of vendors there starting to talk about NVMe over fabrics, ASICs, FPGA implementations, system on chip implementations, software implementations across a variety of stacks. The great thing was NVMe over fabrics was the phrase of the entire show. The challenging thing was probably no two of those solutions interoperated with each other yet. We were still at the running water through the pipes phase, not really checking for leaks and getting to broad adoption. Broad adoption I think comes when we've got a number of vendors, broad interoperability, sort of multi-supplier component availability and those kinds of things that let a number of implementations exist and interoperate because our customers live in a diverse multi-vendor sort of environment. So that's what it will take to go from interesting proof of concept technology which I think is what we're seeing in terms of early customer engagement today to broad-based deployment in both existing fiber channel implementations and also in some next generation data center implementations. Probably beginning next year. So Danny, I talked to a lot of companies out there. Everyone that's involved in the storage industry has been talking about NVMe and NV over fabric for a couple of years now. From a user standpoint, how are they going to help sort this out? What will differentiate kind of the checkbox? Yes, I have something that follows this to, oh wait, this will actually help performance so much better. What works with my environment? Where are the pitfalls and where are the things that are going to help companies? What's going to differentiate in the marketplace? Yeah, so as an engineer, we always get into the speeds and the feeds and the weeds on performance and things like that. And while those are all true, we can talk about fewer and fewer instructions in the network stack, fewer and fewer instructions in the storage stack. We can talk about more efficient silicon implementations, more affinity for multi-processor, multi-core processing environments, more efficient operating system implementations and things like that. But that's just the performance side. The broader benefits come to beginning to move to a more cost effective data center fabric implementation where I'm not managing an orange wire and a blue wire unless that's really what I want. There still are a number of people who want to manage their fiber channel sand and will run NVMe over that. They get the compatibility that they want. They get the policies that they want and the switch behavior that they want and the provisioning model that they want and all of those sorts of things. They'll get that in an NVMe over fabric implementation. A new data center, however, will be able to go, you know what? I'm all in day one on 2,500 gigabit ethernet as my fundamental connection of choice. I'm going 400 gigabit ethernet ports as soon as Andy Bechtelsheim or somebody gives them to me and things like that. And so if that's the data center architecture model that I'm in, that's a fundamental implementation decision that I get to make knowing that I can run an enterprise grade storage protocol over the top of that and the industry's ready. My external storage is ready, my servers are ready and my workloads can get the benefit of that. Okay, so if I just step back for a second, NVMe sounds like a lot of it is what we consider really the backend improving that. NVMe over fabrics helps with some of the front end. From a customer standpoint, what about their application standpoint? Can they work with everything that they have today? Are there things that they're going to want to do to optimize for that? Does the storage industry just take care of it for them? Where do they think about kind of today and future planning from an application standpoint? You know, I think it's a matter of that readiness and what is it going to take. The good news, and this has analogs to the industry change from HDD to SSDs in the first place, the good news is you can make that switch over today and your data management application, your database application, your warehouse, your analytics or whatever, not one line of software changes. NVMe device shows up in the block stack of your favorite operating system and you get lower latency, more IOs in parallel, more CPU back for your application to run because you don't need it in the storage stack anymore. So you get the benefits of that just by changing over to this new protocol. For applications who then want to optimize for this new environment, you can start thinking about having more IOs in flight in parallel. You can start thinking about what happens when those IOs are satisfied more rapidly without as much overhead in interrupt processing and a number of things like that. You can start thinking about what happens when your application goes from 100 microsecond latencies and IOs like with flash devices to 10 microsecond or one microsecond IOs with perhaps with some of these new storage class memory devices that are out there. Those are the benefits that people are going to see when they start thinking about an all NVMe stack, not just being beneficial for existing flash implementations but being fundamentally required and mandatory to get the benefits of storage class memory implementations. So this whole notion of future ready was one of the things that was fundamental in how NVMe was initially designed over 10 years ago and we're starting to see that long-term view pay benefits in the marketplace. Any insight from the customer standpoint as to are there any certain applications or verticals where this is really going to help? I think back to the move to SSDs, it was David Floyer just went around the entire industry and was like database, database, database is where we can have the biggest impact. What's NVMe going to impact? So I think what we always see with these things, first of all, NVMe is probably going to have a very rapid advancement and impact across the industry much more quickly than the transition from HDD to SSD. So we don't have to go through that phase of a little goes a long way. You can largely make the switch and as your ecosystem supports it, as your vendor of choice supports it, you can make that switch and to a large extent have the application be agnostic from that. So that's a really good way to start. The other place is you and I have had this conversation before. If you take out a cocktail napkin and you draw an equation that says time equals money, that's an obvious place where NVMe and NVMe over fabrics benefit someone initially. High-speed analytics, real time, high-frequency trading, a number of things where more efficiency, my ability to do more work per unit time than yours gives me a competitive advantage, makes my algorithms better, exposes my IP in a more advantageous way. Those are wonderful places for these types of emerging technologies to get adopted because the value proposition is just slam dunk simple. Yeah, so running through my head are all the latest buzzwords. Everything at Wikibon when we did our predictions for this year, data's at the center of all of it, but machine learning, AI, heck, blockchain, edge computing, all of these things can definitely be affected by that. Is NVMe going to help all of them? Machine learning, incredible high bandwidth application. Wonderful thing, stream data in, compute on it, get your answers and things like that. Wonderful benefits for a new squeaky clean storage stack to run into. Edge, where oftentimes real time is required, the ability to react to a stimulus and provide a response because of a human safety issue or a risk management issue or what have you. Any place that performance lets you get close, get you at or close to real time is a win and the efficiency of NVMe is a significant advantage in those kinds of environments. So NVMe is largely able to help the industry be ready just at the time that new processing models are coming in, such as machine learning and artificial intelligence, new data center deployment architectures like the Edge come in and the new types of telemetry and algorithms that I may be running there. It's really a technology that's arriving just at the time that the industry needs it. Yeah, I was reading up on some of the blogs on the Dell sites. Jeff Brudeau, Brudeau said, we should expect to see things from 2018. Not expecting you to pre-announce anything, but what should we be looking for from Dell and the Dell family in 2018 when it comes to this space? You know, we've been, we're very bullish on NVMe. We've been pushing very, very hard in the standards community, obviously. We have already shipped NVMe for a series of internal use cases in our storage platforms. So we have confidence in the technology, its readiness, the ability of our software stacks to do what they need to do. We have a robust multi supplier supply chain ready to go so that we can service our customers and provide them the choice in capacities and capabilities and things like that that are required to bet your business and long-term supply assurance for and things like that. So we're seeing the next year or so be the full transition to NVMe and we're ready for it. We've been getting ready for a long time. Now the ecosystem is there and we're predicting very big things in the future. Okay, so, Danny, you've been working on this for like 11 years. You know, give us just a little bit of insight, you know, what you learned, what this group has learned from previous transitions, you know, what's excited you the most and what, give us a little bit of the sausage making. You know, what's been funny about this is, you know, we talk about the initial transition to flash and just getting to the point where, you know, a little goes a long way. That was a three year journey. We started in 2005, we shipped in 2008. We moved from there, you know, flash in arrays as a tier, as a cache, as a, you know, the places where a low latency, high performance media adds value and those kinds of things. Then we saw the industry begin to develop into some server-centric storage solutions. You guys have been at the front of forecasting what that market looks like with software-defined storage. You know, we see that in technologies like ScaleIO and VCN where their ability to start using the media when it's resident in a server became important and suddenly that became, or that began to grow as a peer to the external storage market, another market, you know, San Alternative kind of came along with that. Now we're moving, you know, even further out where, you know, it seems like we used to ask, you know, why not flash? Oh, sorry, why flash? And it will get asked that. Now it's why not flash? Why don't we move there? So what we've seen is a combination of things. As we get more and more efficient, low latency storage protocols, the bottleneck stops being about the network and starts being about something else. As we get more multi-core compute capabilities and Moore's Law continues to tick along, we suddenly have enough compute and enough bandwidth and the next thing to target is the media. As we get faster and faster, more capable media, such as the move to flash and now the move to storage class memory. Again, the bottleneck moves away from the media, maybe back to something else in the stack. As I move, as I advance compute and media and interconnect, suddenly it becomes beneficial for me to rewrite my application or replatform and create an entire new set of applications that exploit the current capabilities of the technologies. And so we are in that rinse, lather, repeat cycle right now on the technology. And for guys like you and me who've been doing this for a while, we've seen this movie before. We know how it ends. It actually doesn't end. There are just new technologies and new bottlenecks and new manifestations of Moore's Law and Ohm's Law and Metcalf's Law that come into play here. All right, so Danny, any final predictions from you on what we should be seeing? What's the next thing you work on that? You could declare victory soon, right? Yeah, so I'm starting to lift my eyes a little bit and we think we see some really good capabilities coming at us from the device physicists in the white coats with the pocket protectors back in the fabs. We're seeing a couple of storage class memories begin to come to market now. You're led by Intel and Micron's 3D crosspoint, but a number of other candidates on the horizon that will take us from this 100 microsecond world to a 10 microsecond world, maybe to a 100 nanosecond world. And you and I will be back here talking about that fairly soon, I predict. Excellent, well, Danny Cobb, always a pleasure to catch up with you. Thanks so much for walking us through all of the pieces. We'll have lots more coverage of this technology and lots more, check out thecube.net. You can see Dell Technologies World and lots of the other shows where we'll be back. Thank you so much for watching theCUBE.