 From London, England, it's theCUBE. Covering Discover 2016 London. Brought to you by Hewlett Packard Enterprise. Now, here's your host, Dave Vellante and Paul Gillis. We're back in London at London Excel. This is theCUBE, the worldwide leader in live tech coverage. This is HPE Discover 2016. Milan Shetty is here. He's the chief technology officer of the data center infrastructure group. Recently promoted. Congratulations. Thank you. Formerly CTO storage group. Now, the entire data center infrastructure group. Fantastic. So, just a few months in, right? Yep, yep. How's it going? It's going great. It's like that this role has got to, if you want to use the industry buzzword, my role now is hyper-converged. Storage and networking put together, might as well say that it's a hyper-converter. It's apropos, right? Exactly, exactly. It's going great. I think so, and as we have been watching the industry, right, and as you have reported many times, compute storage and networking is coming together, right, in many forms and how to get deployed. And it was only a matter of time where even the buying cycles and the hyper-conversion, the converse space has shown. And also internally, organizationally, putting it all together makes a lot of sense from trying to get the change and the integrations done faster, quicker, and in a more agile way. Well, there's no doubt that your organization is converging under Antonio's governance, and so that's good. Where do you see the role of storage now? As a predominant server company, you see storage and server beginning to come together. What happens to this bespoke storage world? Yeah, so I think there's the bespoke storage of the external. Actually, we think that there are three topologies where storage is going to exist. The first topology is the topology which we know externally attached to compute. That's going to be here for a while because there is a need and many use cases where you want to add compute scale independent of the storage scale. So from a use case standpoint, that topology of external storage is going to be around for a while because again, you want to add in so many cases compute independent of storage from a scale standpoint. Second topology is going to be servers with internal storage and you rack them and stack them. So either in pizza box form factor or large form factor and in that case, you are scaling compute and storage in incremental chunks. So that architecture has got some interesting properties. You can't really scale compute and storage independent of each other. So if you need more compute, you're still adding storage. If you need more storage, you're adding compute. But there are a lot of use cases, especially when it comes to data which could be in a Hadoop platform as an example and everything. Perfect fit for that. Large term archival repository, perfect fit for that. That's the second topology. The third topology which is emerging is actually a byproduct of the machine project we had. And we call it the memory-centric computing or memory-driven computing. And the core philosophy there is from an topology standpoint is that in any use cases, the working set which you work with is actually really small in a lot of use cases. If what would the world look like if that working set was in a central pool of memory and there were a bunch of computes which are accessing it simultaneously or synchronously or what have you. The advantage of that architecture is that a lot of applications would actually need less compute because data is in one place, it's preserved, it's persistent and your compute is getting to the data with the memory speed rather than the network moving the data back and forth trying to figure out which computer attaches further. So I think from a storage standpoint and from data standpoint, the three topologies are going to, the one which already exists today, the second of the external storage, the cloud providers made the servers with internal storage as the new player and the memory-centric computing. So storage will actually exist in all three places. So talk a great segue into the machine. This is really the big news coming out today of the conference HPE yesterday demonstrating components of the machine. Yeah. And a lot of tea leaf reading going on right now about what was really important about that demonstration, what should customers take away from what you demonstrated? So great question. So one of the persistent memory and applications benefiting, applications whose dataset can fit in cache. Caches exist today and the cache density is only going to increase. They're at gigabytes today, they're only going to, terabytes you can also put together, but they're expensive. But you can do that, gigabytes and terabytes. What we wanted to show with the machine initially was that when persistent memory comes in and what I mean by persistent memory is that server gets powered off, your data is still there. Server comes back on, you can just read the data rather than having to bring the data from the network and the storage to the compute side. So when we launched the machine program, one of the things we wanted to educate the marketplace was how would the world, and especially the application providers, an application writer, what would a world look like if you never have to be afraid of reboots? The way you would handle things would be very different. Now, we couldn't call it from a marketing standpoint. We're going to make reboot go away. That's not that sexy. From a market standpoint. But here's what containers is doing and also machine is doing, right? In the world of containers, services can be just powered on and off or moved around, the virtualization moved around. Machine is never rebooted. Machine reboot takes minutes, sometimes hours, depending on how long since you last made it because it has to go through all its physical checks and everything. But as you are scaling the applications and everything, you want to be able to just move stuff around, restart the service, take seconds, right? But when you're restarting the seconds, what happens to the data? If the data had the proximity to that container, it might as well be with the compute because that was the working set of the container. And if in that machine, the data is going to be there or in the server. I should not use the machine here or load it. But if in that server, you never have to go get the data from the storage. Your working set is always in the cache. You're keeping the second copy for protection reasons and everything on the external storage. How you write applications, how you deploy applications and how you scale applications dramatically changes. You could just completely dramatically changes and also that has a big cost benefit. You may not need as much cores and you may not need as much external storage. But avoiding reboots or shortening reboot time doesn't seem to be a very compelling use case. I was using it as one example. When you are scaling at large, it is one of the side effects. That's why I was mentioning that it's not the containers that makes it easier. No, it's the, when you are scaling services and when you have thousands and thousands of machine, you're trying to do a firmware upgrade as an example and you have to take the machine down. What happens to the services? There are large outages or large scale deployments actually don't want to change their infrastructure. Don't want to add capacity. Don't add, because the limits of the, limits of the cloud infrastructure breaks very fast. What we're saying is that by doing memory centric and one of the persistent memory, you actually need less infrastructure. That's the first benefit. Because your working set is right there, right close to the data. So the first benefit is really the working set is closer so you need less infrastructure to do the same amount of job. So that's a big win. Faster. And the second win is you can actually service your, you can decouple the servicing of the hardware from servicing of your application. And that's a big benefit for a large, large set of service providers. So you're moving data around less, swapping it between storage and memory. It's all in one big shared memory pool. How is this enabling the kind of dramatic performance improvements that HPE was talking about yesterday? Absolutely. So if you look at the entire working set of, so I'll just pick traditional applications like which are built on Oracle and SQL and everything. Are just Oracle or SAP instances you will see is like handful of terabytes. After you do three or four terabytes. That's all is now. That's not going to need required network bandwidth at all. It's sitting right next to compute. So your entire Oracle query is going to run at memory speed. Because the eight terabyte, 10 terabyte Oracle databases are considered big size databases. And now you can just do it in one blade. So can you talk about practically how this is going to show up in products? HPE was talking about 2018, 2019. Are we going to see a device called the machine? Or are we going to see this technology parade the whole product line? It's the ladder. There is not going to be a thing called the machine. And then everybody writes to the machine. Machine will be the concept and the technology which will change the way, the memory centric way on how the applications are going to be deployed and sped up. So the first place where the technology will show is actually going to be in part of my whole job in storage. So if you think about it, pre-par working set, when you look at the storage systems architecture, you have the, if you just look at the external storage and if you open up the external storage, it actually looks like compute. It's got CPU, it's got memory, it's got storage and it does a lot of caching and everything. So now with persistent memory and big persistent memory out there on the motherboard itself, your entire work- No caching. No caching. No flash. No flash and everything. You will actually need less of the north-south traffic. You need, in case of a three-part, less north-south traffic, which means it can host more virtual machines. Which means you can now host more containers in the same footprint or perhaps even lesser. And that medium is memrister? So it could be, it doesn't have to be memrister. It could be battery-backed dims because they have the similar properties as memrister. It could be battery, I think initially, it will be battery-backed dims and 3D cross-point. Right, okay, so if fundamentally you're talking about Milan collapsing this storage hierarchy, which is caches and DRAMs and non-volatile RAMs and flash and spinning disk gets collapsed. Now that doesn't necessarily say sheep and deep goes away, maybe tape and, okay, fine. What about cost? Yeah, so the persistent memory of the 3D cross-point have reached a place where their cost is very comparable to what we get today as DRAMs. And it will only go down because the volumes will be there. The other use case I was going to mention in addition to the storage use case is IoT. Are you going to move all the data from all the sensors to an public cloud or your IT as a cloud? Or you're going to do some computing closer too where you collected the data. And even in the IoT set of things, right, so you can envision smaller servers, self-contained, self-packaged, then they will sit closer to the edge. They will do all their computing, they did the entire database instances and everything they did right there. And streaming analytics and then feeding subsets of that data up to the server. Yeah, and only the change logs or change delta will be sent back to the central IT shop, but all of the compute work and everything is done there because there is not enough bandwidth available today to get all the IoT sensors data to a central data center, process them and send them back. A, you don't have the time and you don't have the bandwidth. Time isn't the real time decision making and everything. So that is where this is going to be revolutionary. IBM is putting a lot of its research effort into quantum computing right now. HP has been focusing on the memory technology. Are these competitors with each other? No, they're not. They're different approaches. They're different approaches. And what it may mean by that is that we believe that the center of the universe is going to be not compute. And it's kind of hard for an compute company center of the universe is going to be data and memory. And compute will attach to that pool of memory and storage will attach to the pool of memory for keeping second copy or longer copy, longer archival and everything. But the HPE's approach and thanks to our machine project and also all of the data we got from the IoT use cases and everything we talked to the customers around that is the center of the universe for the next decade is going to be memory, not compute. So flash in that scenario becomes some servient to that persistent memory. Yeah, flash would be the backup copy of the persistent memory or longer copy. So I'm going to come back to cost. You said today persistent memories are competitive with DRAM. Okay, what about competitive with flash? I mean, this has a huge advantage. Does it not? That's right. You could use that if you go back in history, that was the same case between flash and disk. The flash economic change because smartphones happened and the volume existed. Right now there are only two DRAM suppliers really. And they're both South Korean companies and the market is captured by two DRAM suppliers and the material cost versus the cost we pay are not, the economics is not following the Adam Smith theory. So just like when smartphones came in there were going to be billions of smartphones and rather than having a smartphone with hard disk drive, the smartphones got flash in. The pivotal, the tipping point of this is going to be IoT because IoT cannot afford flashes too slow in some instances. The volume is going to come from IoT and so when you have millions of sensors and each one of has got small chip which really is a memory chip. That changes the volume for just like what smartphones did to disk and flash industry, the thesis is that the IoT is going to do the memory and the flash industry. And memory, volatile memory plus flash is not going to be in your scenario, is not going to be the paradigm because non-volatile memory will replace the volatile memory, the DRAM. That's right because the cost of battery. And then that volume will allow it to cross with flash. That's right, that's right. Well at least it's a plausible scenario. We'll see if the volume's... It's a long way to go. There's a long way to go but just like the smartphones just showed up and one day disk and flash transition happened. Less than a decade. Less than a decade it happened and I think that we'll start showing up persistent memory use cases in the storage arrays. Then we'll show up in the server product lines and once people realize that actually you know what? And then it's all about packaging, right? Then the IoT devices can just package the non-volatile memory, take a compute and just pack up some non-volatile memory and only send change logs to the central IT shop maybe on the flash devices because it changes in a hurry. That's why our strategy is memory centric, not CPU centric. We agree, the data stays at the edge, 95% of it anyway. Exactly, exactly, exactly. And if we follow that trail, the data says that the edge and most of the data is going to be captured by IoT and then IoT is going to do to memory and flash what smartphones did to flash and disks. That's why we're focused on that. All right Mila, good discussion, we got to leave it there. Thanks very much for coming to theCUBE and sharing your insights. Thank you, thanks for having me. You're welcome. All right, keep it right there everybody. Paul and I will be back with our next guest right after this short break.