 Welcome to Cube Conversations. Cube Conversations is a new series by Silicon Angle Wikibon where we bring people into our studios. We're live here from Marlboro, Massachusetts. Greg Shears here is the Vice President of Server and Storage Strategy at Broadcom. And we're going to talk about trends in storage and networking. Greg, welcome back to theCUBE. Thank you very much. It's great to be here. Yeah, so we see you at all the shows. You know, we travel around, we go to the events and it's good to have you here. Tell us what's new in Broadcom's world. Boy, things never stay the same for very long. We have a lot of new things going on, both on the switching infrastructure side. We've officially released our Trident 2, our highest port count switch. Very exciting, used in an awful lot of the switch infrastructure that's out there these days. We've also done things on the very opposite end too in that we have, we just announced a few weeks ago, 5725 integrated BMC one gig and BMC controller, which you think, boy, it's kind of an odd thing for Broadcom to get into, is kind of the baseboard management controller. But this really plays into, I think, a lot of the topics that we want to talk to today. And it's really about the cloud. So we kept being approached by a lot of the cloud vendors, infrastructure folks saying, you know, we really want a very lightweight management controller. You guys have kind of the industry-leading one gig controller, a lot of the 10 gig controllers as well. What can we do to simplify and have an open management controller for kind of the, I'll call it the lower end or the hyperscale kind of environment, not locked into any given vendor using open standards, DMTF based smash kind of standards. So we have things on the management controller side, lots of 10 gig software enhancements that were constantly rolling out and really looking at a lot of the cloud applications that I think we'll talk about today too. Yeah, definitely. So if you look at the trends in the last 15, 20 years, you've seen functions sort of move out of the host's server into the sand and you're seeing the pendulum swing back and a lot of things driving that. There's flash, there's convergence, you talked about cloud. So what are you seeing in terms of the bigger picture, storage trends, and then we'll get in deep? Boy, a tremendous amount of movement in storage and that if we look at the cloud, the cloud is converged. We hear about in the classic enterprise and an awful lot of the time, we hear people talk about storage convergence. Well, the cloud, I'd like to say the cloud was really born converged. People don't have separate storage appliances that they put in that are storage only. These are really servers with a global file system or an object-based file system. And that's how people do storage in the hyperscaler cloud type environments. And so we see that trend growing. The amount of storage is certainly growing at a huge rate within these hyperscaler public cloud type environments. And so we see some of those same trends moving over away from just say the hyperscale or large public cloud mentors into even private cloud and some of the enterprise as well. Stu, you've written a lot about that factor right there, the sort of hyperscale bleeding into the enterprise. Yeah, so Greg, it's interesting, we talk about the components breaking down. If you look at what Facebook is doing with open compute, we said, can we just completely disaggregate the pieces? I remember years ago when Facebook's photo repository was growing, they used to use traditional filers from a big name vendor and at reach a point they were growing too fast and the cost just wasn't there that they had to change the architecture. But I think those of us watching, it's not gonna break down to everybody buying their own chips and pieces together. There has to be some balance there. And we heard from Amazon last week that they don't think it should be some just giant pool. We say Facebook built five configurations and everything can run off that. We heard from Amazon, they actually hyperspecialize what they're doing and have more and more configurations because when they scale each application, they wanna configure that and have those pieces in place. So how do you see those trends playing out? Well, it's so true. Stu, we see more and more the whole notion of this hyperspecialization. I mean, earlier this year, we heard about Baidu coming out with kind of their storage server that was really a multi terabyte kind of enclosure and a 2U form factor, not really meant for speed, but meant for the term that we've kind of coined after the fact is more cold storage. We hear Facebook talking about that as well. Amazon Glacier was another one, right? Yeah, well, Facebook had talked about, oh goodness, they talked about just in terms of uploading pictures, the state is a little bit old now, but really only about a quarter old. They do somewhere in the neighborhood of a petabyte of raw storage just in pictures a day and then they store those in different image formats and so by the time they store the different image formats and they index it, they're looking at about north of 10 petabytes a day in terms of their storage growth and you think, my gosh, what do you do? How do you keep up with that? And I think that's, to your point, you can't have a traditional filer or a traditional storage array keep up with that. So you have to do some very specialized things in order to keep up with that kind of traffic. So you're seeing the spinning disc bottleneck being addressed by Flash, the cold storage, you don't care so much about latencies and the like. So is network, is the network becoming the bottleneck and how is the industry dealing with that? That's a great point, because here to four, a lot of one gig was used in the enterprise and one gig networking was fine for messaging, but now that we're talking about richer and richer content where this content is, as we talked about petabytes worth of data, there's no way you can keep up with that with one gig network. So much of the hyperscale environment has already moved to 10 gig right to the edge. Some of the public cloud infrastructures is rapidly 40 gig, very, very fast network speeds, and a lot of that's to basically globalize their storage pools. So rather than having one giant array that had petabytes worth of data behind that array, now you distribute those petabytes and you have the data fall wherever you can fit it in specialized servers that have lots of storage attached to them. But in order to get at that data now, you end up having very fat pipes, which really has a whole trickle down in terms of how you build your networks. You no longer have aggregation level switches that you have one gig at the edge and maybe 10 gig in your backbones, you have 10 gig all the way to the edge. So your backbones, you have a full cross-sectional bandwidth that's 10 gig everywhere and moving to 40 gig. So I want to understand the premise on this whole notion of data locality. You're saying that increasingly customers will sort of put the data wherever it's most convenient and then you'll architect the network globally, but you're still gonna have fat pipes to get to it because you still gotta move data or function or metadata to and from that location, is that right? That's absolutely it. One of the first premises of the big data environment was actually to move compute to where the data lived because the data was just so massive, but that was predicated on very small pipes. So one gig kind of networking, inexpensive networks. Now what we're seeing is that the data is so massive in some cases. Google has talked about the amount of flat data that they capture on a daily basis being measured in petabytes in the North American market alone. So what ends up happening is you end up having to move that data around in order to make effective use of it. So the idea is your pipes get much fatter to where the idea is to still try and move the compute to where the data is, but many of the servers can't hold as much data as needed to process. So your network is really your universal transport to move that data around, especially the metadata, as you are capturing results from that initial data and storing it elsewhere. You're using the network as your transport. So if I can leave much of the data where it is, act on the metadata, maybe the metadata sitting in fast flash or even in memory, that just changes the whole way in which we look at. Yeah, so, Craig, I wonder if I could poke at that a bit because I can hear Wikibon CTO and co-founder David Fleur yelling in the back of my head, even with fat pipes, the laws of physics rule. And if I have a large amount of data and some of it's here and some of it's there, transporting that data takes a long time. So does it all need to come together? So I think about Amazon has direct connect. So I want to have a fat pipe sitting right between my location, maybe it's at Equinex or something, going to Amazon. Do I want a colo where I can have my data that I own it, but I have a cloud that basically I can put just a fat pipe over a wall because if we're talking about miles between my data, even if I have a fat pipe, it's not going to get from here to there. No, it's a great point. And one of my favorite stories about that in terms of fat pipe versus overall bandwidth is that many of the huge mega data centers, their fastest data transfer are things like FedEx where they'll take disks that they'll go ahead and do local backup and they'll ship it from one site to the next because they get the best bandwidth that way. And you think, how can that be? But yet even 40 gigabit, 100 gigabit, when you're transferring pedabytes of data, it takes a huge amount of time. That's the technical term for that, those who watch Wikibon SiliconANGLE, the Chevy Truck Access Method. That's it. I love it. So I mean, it's still, locality is still important, but there are some applications and this is the part that we see more and more. The applications are being tailored to what the infrastructure can provide. And so in environments where you truly have data that's so large, you have to make sure that your application is distributed so that you are not moving the data as much. Yeah. So Greg, while we've got you here and we're on this topic, we always talk about, what's the adoption of some of these trends? So in the enterprise we still see at the edge of the server level, it's still a lot of one gig. I think still definitely over 50% one gig trend and 60, 70% maybe, 10 gig and 40 gig starting to come in some of the core. Whereas if I talk about kind of the internet, it's got 100 gig on some of the backbone and the cloud guys have 10 gig pervasive, 40 and 100 gig starting to come out. What are you seeing and what do you expect to see over the next year or two there? What's holding us back? Boy, a lot of interesting trends just in general but I completely concur with you. I mean, we're focused a lot on looking at the enterprise rack and tower where the penetration of one gig is unfortunately, it's about 90 some odd percent, really only about eight to 10% migration or penetration of 10 gig in that classic enterprise rack and tower. We are starting to see signs of that moving though and if we talked about blades, well, blades are predominantly 10 gig today, 85, 90% 10 gig. But we're really looking at the trend of how to get more 10 gig in that rack and tower which is the bulk of the overall enterprise. Tenchi-based T is starting to make more of an impact in that market. That market tends to be dominated by end of row kind of switching and in that kind of environment, just a simple low cost copper DAC cable for 10 gig, it won't reach to the end of row. So Tenchi-based T makes a lot of sense. It does let you upgrade your servers and your switches independently. And now we see from virtually all of the switch vendors, 10 gig Tenchi-based T switches being offered that are pretty competitive and much, much lower power than first, second, even third generation capabilities. So how are the cloud guys affecting this adoption? I mean, you see, if you add up Google and Microsoft and Facebook, it's starting to be more percentage of the overall ports being shipped. I heard it might even be 20% of overall port shipments today, where's their adoption driving things? Well, I think in general, the whole of the cloud or mega data centers, what they do is they give the enterprise, I think, a vision for where they can go from both the economy of scale standpoint and just looking at applications through a new lens. The enterprise still moves quite a bit slower. Part of that's because their scale is much, much lower. When you're putting in an entirely new data center, the way Facebook and some of these other folks do, or deploy tens of racks at a time to deploy an entire application on this new hardware, the classic enterprise typically doesn't do things like that. They typically replace their hardware at the end of a three year or five year lease. And so just the adoption cycles are very different. But having said that, I think what we are seeing is is we're seeing that the costs, some of the classic OEMs have really reduced their pricing strategies on higher speed networking gear and NICs, and I think that can do nothing but help the adoption rates. And plus, I think as virtualization gets more of a foothold, now it's no longer sort of a nice to have technology, it's a mandatory technology, and the scale of it keeps getting larger and larger. And 10 gig is a natural, 10 gig and even beyond is required when you put tens of physical servers on one virtual server platform. So you're watching CUBE Conversations, we're here with Greg Shearer of Broadcom, talking about storage and networking trends. Greg, we were talking about some of the trends around convergence and cloud. SDN, hot buzzwords, you hear it when you go to VMworld and now all over, all kinds of action going on. What's your take on SDN? Does it increase the need, for instance, for specialized appliances? Does it decrease the need? And we're just gonna have commodity components running everywhere, what's your point of view on that? Boy, I mean, SDN is certainly all the rage. You can't open up a trade rag, excuse me, for those of you that publish those. But SDN is everywhere in terms of all of its nomenclature. You know, from Broadcom's standpoint, we think SDN is wonderful. And I do think that it's gonna drive commoditization. Broadcom's not afraid of commoditization, there's many companies that hear that C word and it's a four letter word in their vocabulary. For Broadcom, we're all about volume. So for us, having something that moves towards commodity and being able to use specialized software, but on commodity kind of appliances, server appliances, we've seen this happen with storage in general, in the whole cloud environment. That's really the basis now for storage within the cloud in mega data centers. It's just commodity servers, very specialized software, so that the software is what gives the storage its unique value, global file systems, global object spaces, things like that. And we see that trend increasing and SDN is really gonna do for the whole of the switching infrastructure. And I think even beyond switching infrastructure, it really allows whole applications to be built around this whole concept of building the entire software defined application in both looking at the network and all the compute components as individual components that can be plugged in as needed at the lowest possible cost points. So talking about applications, one of the new emerging applications is of course, big data, you guys are doing some work in big data. You have a big data lab, talk about that a little bit. You bet, so we're very excited because you know, Broadcom is traditionally, you know, and I'll say not just Broadcom, but most networking vendors, we've really focused on microbenchmarks, so things like RFC 2544 to look at what our latency characteristics were or the number of packets per second that we can do at a given network packet size. And those are the things that you'll see published from most folks. What we're really focusing on now, especially when we look at big data, is building out an ecosystem and then testing in the real world and finding out what are the things that we can accelerate when we get closer to that application. Big data being, you know, one of the most important things going on now. But big data is really a conglomeration of there's a compute aspect to big data. There's, you know, a networking aspect in terms of acquiring the data and then sending our results over the network to a different home and then extracting the data out of those results and exporting them maybe to some other kind of SQL or no SQL kind of database. So there's multiple aspects. And so what we've done is we've built up a big data lab to go ahead and really simulate those. So running that in real world environments with tunable parameters to find out what made the biggest difference. What difference does latency have on the overall workload? What difference does bandwidth have on the workload? What difference and what are the things that we could do to accelerate those? So in addition to kind of the big data lab, we've also been looking at sort of web 2.0 workloads, Memcache D, looking at ways to accelerate the application. So getting away from kind of the micro benchmarks to look at how fast a given operation can be done, like how fast we can transport a packet from point A to point B, versus let's look at it from the application standpoint and look at what the response time is for a particular query and what we can do to influence that and how many of those per second we can do from an application perspective as opposed to just looking at it from the micro benchmarks. So Greg, it's so interesting to hear you say that because I think back before we were optimizing, as you said, TCP or iSCSI or some of these things because it was just Microsoft, and I had a billion applications that sat on Microsoft and they all had different requirements. It sounds today there's a few really key applications that are gonna be done at such huge scale that you could focus, because if I think back, 10 years ago, there's no way any company could have said, oh, we're gonna optimize for an application because there's so many of them. So is that what you're seeing then? Are we gonna have, does Hadoop own half the world's storage in five years from now? As I think that was Eric Baldishwire from, made that quote. So where do you see applications going in the future? Well, I think that's such an important aspect, especially as we look at where the world's going in scale. Hadoop has revolutionized the way we store data, the locality issues of pushing compute to where the data is, where possible. But even down to things like Memcache D, to where, and if you blur your eyes a little bit, Memcache D is nothing more than a distributed application to basically look up billions or trillions of objects through a key pair lookup. So you think, well, if we don't call it Memcache D, but we look at it in terms of other applications, what can hardware do to help accelerate those lookups? Because a lot of those lookups are memory-based lookups to where you look up in one table, you hash a key, and then that takes you to a secondary table and another table and so forth. Well, hardware can be quite good at some of those operations, hashing specifically, CRC calculations. So that's the level that we're looking at, but to your point, it's fascinating when you start looking at this from the standpoint of these applications or applications that are derived from Hadoop, Haystack is another application. If you figure out how to accelerate things in Hadoop or things in Memcache, they naturally follow to a whole suite of other capabilities as well. So I think the world is getting more diverse in some respects, but more commonality between the different factions and diversities. So Greg, I love when you come on theCUBE because you could talk about a lot of different things. You could talk about the host side, you could talk about controllers, switches, storage, networking, obviously. Tie it all together for us. You run strategy for server and storage. You mentioned earlier, you welcome commodity. So from a strategy standpoint, how are you guys taking the company sort of beyond those traditional components that are becoming more commodity-like and moving into more of a business value discussion? Oh, you bet, thanks, because we really do see a tremendous amount of change in the data center. And I think 10 years ago, we were almost exclusively focused on hardware-based OEMs and hardware-based distribution methods of Broadcom capabilities. Now, many of our partners are really software partners. Folks that are looking at adding value, both in SDN and Hadoop, general big data, a lot of the applications themselves, and figuring out how to make it easy for those to proliferate through the entire data center. So having a seat in the host as well as a seat in the network I think gives us a very unique position to look at sort of globally looking at the applications and figuring out how to not only accelerate them but make it easier to deploy. Many of these things become buzzwords in the industry. SDN is kind of the current buzzword in the industry. But I think we're also looking at ways to sort of demystify that and make it more practical, take some of the mysterious aspects of it and through our SDK software development kits and through software that we work with partners, figure out ways to show the value proposition and push that out into the marketplace. Well, Greg, thanks very much for coming on to CUBE Conversations. You've been watching on slash Silicon angle. Thanks for watching everybody. We'll see you next time.